AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
Human Preference aligned Tabular Similarity
Theory
- Current tabular embedding methods are primarily optimized for prediction tasks, not for human preference alignment in similarity search.
- Standard evaluation metrics do not adequately assess the trustworthiness of embeddings for similarity-based applications.
- The proposed workflow for human preference ranking includes embedding, anchor selection, and user engagement.
- Human preference alignment is crucial for ensuring that retrieved items in similarity searches are relevant to domain experts.
Read more
Human Preference aligned Tabular Similarity
Summary
This paper addresses the limitations of existing tabular embeddings used for similarity search in industrial applications, particularly in Product Lifecycle Management (PLM) systems. While current embedding methods are optimized for prediction tasks, they do not effectively align with human preferences for similarity rankings. The authors argue that existing evaluation metrics fail to capture the subjective nature of similarity as perceived by domain experts. They propose a new evaluation procedure that incorporates human preference alignment, emphasizing the need for a more trustworthy approach to assess the relevance of retrieved items in similarity searches. The paper outlines a workflow for human preference ranking that includes embedding tabular records, selecting anchor records, and engaging users in the evaluation process. This approach aims to enhance the trustworthiness of AI systems by ensuring that the retrieved nearest neighbors are meaningful to users, thereby addressing the critical gap in evaluating task-agnostic embeddings for similarity-based use cases.
Methodology
The authors propose a three-step workflow for human preference ranking of embedding similarity in tabular data, which includes: 1) Ingestion of tabular records into a vector store using a selected embedding model, 2) Selection of an anchor record based on predefined probability measures, and 3) Continuous user engagement to refine retrieval quality.
Results
The paper illustrates the inadequacies of current evaluation methods through a PLM use case, demonstrating that embeddings optimized for classification may not yield meaningful results for similarity search. The proposed human preference aligned approach aims to bridge this gap and enhance the relevance of retrieved items.
Implications
This work has significant implications for industrial applications that rely on similarity search in tabular data, suggesting that incorporating human preferences can improve user trust and system performance. It encourages the adoption of more comprehensive evaluation standards in AI systems.
Multiclass Classification without Labels via Posterior Simplex Geometry
Theory
- Introduces a framework for multiclass classification without labels, extending the CWoLa principle.
- Demonstrates that the Bayes-optimal mixture classifier can be represented in a (K-1)-simplex geometry.
- Proposes two methods for recovering latent class structures: post-hoc simplex fitting and a bottleneck architecture.
- Validates the approach on multiple datasets, achieving significant performance improvements in label-scarce scenarios.
Read more
Multiclass Classification without Labels via Posterior Simplex Geometry
Summary
This paper addresses the challenge of multiclass classification in scenarios where reliable instance-level labels are unavailable. The authors extend the concept of Classification without Labels (CWoLa) from binary to multiclass settings, demonstrating that a classifier can learn to distinguish between different mixtures of classes based solely on their identities. The core idea is that the Bayes-optimal mixture classifier maps data points into a (K-1)-simplex in mixture-posterior space, where K represents the number of latent classes. The vertices of this simplex correspond to the latent classes, allowing for the recovery of class structures and proportions without direct observation of labels or mixing proportions. The authors propose two main approaches for latent class recovery: post-hoc simplex fitting and a bottleneck architecture. They validate their methods on several datasets, including MNIST, CIFAR-10, and Galaxy10 DECaLS, showing that their approach can effectively recover latent classes and their proportions, thereby bridging the gap between weakly supervised and fully supervised learning.
Methodology
The authors train a classifier to predict mixture identities from unlabeled data and utilize the geometry of the resulting posterior simplex to recover latent class structures. They employ two main techniques: post-hoc fitting of the simplex to the learned posterior and a bottleneck architecture that directly maps inputs to the simplex vertices.
Results
The experiments conducted on MNIST, CIFAR-10, and Galaxy10 DECaLS demonstrate that the proposed methods can successfully recover latent classes and their proportions from unlabeled mixtures, achieving performance that narrows the gap between weakly and fully supervised learning.
Implications
This work provides a scalable and mathematically grounded approach for multiclass discovery in domains where labels are scarce, potentially benefiting fields such as scientific research and clinical applications where expert annotations are difficult to obtain.
Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB
Optimization
Efficient ML
Theory
- Kairos addresses the item cold-start problem in news recommendation systems.
- The framework utilizes a Cholesky-based approach for numerical stability in LinUCB.
- Integration of Matryoshka Representation Learning enhances inference efficiency.
- Empirical results show a significant efficiency gain while maintaining ranking precision.
Read more
Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB
Summary
The paper presents Project Kairos, a framework designed to enhance news recommendation systems in environments characterized by item cold-start challenges. Traditional deep learning models require extensive interaction data, which is often unavailable in the fast-paced news domain where articles have a short Time-to-Live (TTL). Kairos addresses this issue by employing a contextual online learning approach using the Linear Upper Confidence Bound (LinUCB) algorithm. A key innovation is the replacement of the error-prone Sherman-Morrison matrix inversion with direct rank-1 updates of Cholesky factors, ensuring numerical stability and preserving the positive definiteness of the covariance matrix even under ill-conditioned data scenarios. Additionally, the integration of Matryoshka Representation Learning (MRL) allows for adaptive inference, reducing latency while maintaining ranking precision. Empirical evaluations using data from the Tagesschau API demonstrate that Kairos achieves a 4.85-fold efficiency gain in candidate generation without significantly compromising the quality of recommendations. This framework provides a robust solution for high-performance recommendation systems in data-scarce environments.
Methodology
The methodology involves transitioning from passive data-driven modeling to active online learning using the LinUCB algorithm, which estimates recommendation payoffs through contextual embeddings. The framework replaces traditional matrix inversion techniques with Cholesky factor updates to ensure numerical stability. Additionally, it employs Matryoshka Representation Learning for efficient candidate generation.
Results
Empirical evaluations indicate that Kairos achieves a 4.85-fold increase in efficiency for candidate generation while preserving ranking precision, demonstrating its effectiveness in handling the cold-start problem in news recommendation.
Implications
The findings suggest that Kairos can significantly improve the performance of recommendation systems in environments with limited data, making it applicable to various domains where rapid content turnover occurs, such as news and social media.
Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions
NLP
Large Language Models
Time Series
- Recast framework enables proactive forecasting of safety risks in multi-turn interactions.
- It retrieves evidence from both short-term and long-term contexts to model risk evolution.
- Achieves 88.3% accuracy in predicting future safety failures with a lead time of 2.41 turns.
- Addresses the limitations of existing reactive safety measures by focusing on trajectory-level risks.
Read more
Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions
Summary
The paper addresses the challenge of forecasting safety risks in multi-turn interactions involving large language models (LLMs). As LLMs evolve into autonomous agents, traditional pointwise risk assessments are insufficient for predicting how risks develop over time. The authors introduce 'Recast', a framework that shifts the focus from reactive detection of safety violations to proactive trajectory-level risk forecasting. Recast retrieves risk-relevant evidence from both short-term dialogues and long-term historical contexts, modeling the evolution of risks through a structured approach. It employs a causal temporal encoder to learn latent patterns of risk evolution and predict future risk emergence. Experimental results demonstrate that Recast successfully predicts 88.3% of future safety failures with an average lead time of 2.41 turns, while maintaining a false alarm rate of 12.3%. This approach highlights the importance of understanding the dynamics of multi-turn interactions in ensuring the safety of LLMs.
Methodology
The Recast framework consists of three main components: (1) Risk-relevant historical evidence retrieval, which captures temporal dependencies from dialogue progression and historical context; (2) Compositional risk state and transition modeling, which structures risk states and models their evolution; and (3) Future risk distribution forecasting using a causal temporal encoder to predict future risk emergence patterns.
Results
Recast predicts 88.3% of future safety failures with an average lead time of 2.41 turns and maintains a false alarm rate of 12.3%. The framework achieves a negative log-likelihood of 0.45 and a mean absolute error of 0.31 for first-risk emergence time forecasting across seven risk categories.
Implications
The findings suggest that proactive risk forecasting can significantly enhance the safety of LLMs in multi-turn interactions, paving the way for more robust and secure autonomous systems. This approach could be applied in various domains where LLMs are used, such as customer service, healthcare, and autonomous decision-making systems.
Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
Reinforcement Learning
- Introduces the Reinformed Dreamer algorithm, which addresses limitations in the Informed Dreamer.
- Proposes a new learning objective using a privileged variational encoder for better representation of privileged information.
- Demonstrates improved performance and convergence speed over existing asymmetric reinforcement learning approaches.
- Maintains a single world model structure to simplify the imagination process while enhancing learning efficiency.
Read more
Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
Summary
The paper introduces the Reinformed Dreamer, a novel algorithm in the field of model-based reinforcement learning that enhances the learning process by utilizing latent guidance. The authors identify limitations in the Informed Dreamer, an existing asymmetric model-based algorithm, particularly regarding the learning of privileged information representations. They propose a new learning objective that employs a privileged variational encoder alongside an unprivileged encoder, allowing for effective representation learning without complicating the imagination process. The Reinformed Dreamer demonstrates improved convergence speed and performance across various benchmarks compared to previous methods, showcasing the benefits of leveraging additional information during training to optimize representation learning in partially observable environments.
Methodology
The authors developed the Reinformed Dreamer by integrating a privileged variational encoder with an unprivileged encoder within a single world model. This approach allows for effective representation learning of both privileged and unprivileged information. The algorithm was tested across various benchmarks to evaluate its performance and convergence speed compared to existing methods.
Results
The Reinformed Dreamer outperformed or matched the performance of previous algorithms, including the Informed Dreamer and Scaffolder, in terms of both convergence speed and final performance across several benchmark environments. This indicates that the new learning objective and model structure effectively enhance the learning process in reinforcement learning.
Implications
The findings suggest that incorporating additional supervisory information during training can significantly improve the efficiency and effectiveness of reinforcement learning algorithms. This has potential applications in various domains where decision-making under uncertainty is critical, such as robotics, autonomous systems, and complex simulations.
Thinking Under Uncertainty: Evidence Use and Information-Seeking in Language Models
Large Language Models
NLP
Theory
- Inference-time thinking enhances value-guided actions in language models.
- Thinking does not promote information-seeking behaviors as measured by UCB-like or Thompson-like exploration.
- Longer thinking traces are associated with information-imbalanced histories.
- Reported confidence becomes more sensitive to decision difficulty and task evidence.
Read more
Thinking Under Uncertainty: Evidence Use and Information-Seeking in Language Models
Summary
This paper investigates how inference-time thinking affects the performance of large language models (LLMs) by distinguishing between effective evidence use and information-seeking behaviors. The authors conducted experiments using ten open-weight models in a two-armed bandit task, measuring action preference, thinking length, and reported confidence under matched uncertainty. The results showed that thinking improved value-guided actions and reduced choice noise, but did not promote information-seeking behaviors characterized by UCB-like or Thompson-like exploration. Additionally, the length of thinking was influenced by the history of observations, and reported confidence became more sensitive to decision difficulty. The findings suggest that while thinking enhances decision-making based on current evidence, it does not necessarily lead to a shift towards seeking additional information. The study contributes to understanding the cognitive processes underlying LLMs and their decision-making capabilities under uncertainty.
Methodology
The authors adapted a horizon-based two-armed bandit paradigm into a multi-turn dialogue format, where models made decisions based on controlled warm-up histories of rewards from two slot machines. They measured action preferences, thinking lengths, and reported confidence across different observation conditions. A cognitive model was employed to separate value-guided actions from exploration behaviors.
Results
The study found that thinking strengthened value-guided actions while not supporting a shift towards more information-seeking policies. Greater thinking lengths were observed in information-imbalanced conditions, and reported confidence was more closely aligned with decision difficulty and task evidence.
Implications
The findings have implications for the design and training of language models, suggesting that enhancing metacognitive processes could improve decision-making under uncertainty. Understanding how LLMs utilize evidence and seek information can inform future developments in AI systems that require robust reasoning capabilities.
High-Order Markov Blanket Discovery via a k-Order Relaxation of the Faithfulness Assumption
Graph Learning
Theory
- Proposes a k-order relaxation of the faithfulness assumption to capture higher-order dependencies.
- Introduces the k-order Markov blanket (kOMB) algorithm for discovering Markov blankets.
- Empirical results show kOMB's effectiveness in recovering Markov blankets under violations of faithfulness.
- Demonstrates improved performance over existing methods on benchmark datasets.
Read more
High-Order Markov Blanket Discovery via a k-Order Relaxation of the Faithfulness Assumption
Summary
This paper addresses the challenge of discovering the graphical Markov blanket (MB) of a variable from data, which is crucial for applications in Bayesian networks, causal discovery, and feature selection. Traditional methods rely heavily on the faithfulness assumption, which states that conditional independencies in the distribution imply the same separations in the graphical structure. However, this assumption can be violated by higher-order dependencies, such as XOR and parity relations, leading to spurious dependencies in finite samples. To tackle this issue, the authors propose a k-order relaxation of the faithfulness assumption that accommodates parity-type relationships among k+2 variables. They introduce a proof-of-concept algorithm called k-order Markov blanket (kOMB) that utilizes this relaxation for MB discovery. The empirical results demonstrate that kOMB effectively recovers the MB of a variable even under true and empirical violations of faithfulness, outperforming existing constraint-based methods on benchmark datasets. The findings suggest that exploring relationships among four variables at a time can yield significant improvements in MB discovery.
Methodology
The authors developed a k-order relaxation of the faithfulness assumption to account for higher-order dependencies. They implemented the k-order Markov blanket (kOMB) algorithm, which leverages this relaxation to discover the Markov blanket of a variable from sample data. The methodology includes theoretical proofs and empirical evaluations on benchmark datasets to validate the algorithm's performance.
Results
The kOMB algorithm successfully recovered the Markov blanket of a variable in scenarios where traditional methods failed due to violations of the faithfulness assumption. The empirical evaluations indicated that kOMB outperformed existing constraint-based methods, particularly in cases with limited sample sizes, demonstrating its robustness against both true and empirical faithfulness violations.
Implications
The findings of this research have significant implications for various fields, including causal discovery, feature selection, and structure learning in Bayesian networks. The k-order relaxation approach could lead to more accurate models in scenarios where traditional assumptions do not hold, enhancing the reliability of insights drawn from data.
The Art of Not Forgetting A Local Learning Architecture for Continual Learning
NLP
Theory
Efficient ML
- CMP architecture combines sparse representations, competitive memory, and local learning to address catastrophic forgetting.
- Empirical evaluations show CMP significantly reduces backward transfer compared to a Transformer with EWC.
- The study emphasizes the importance of reporting both positive and negative results in scientific research.
- CMP's movement-based plasticity mechanism allows for parameter updates without relying on gradient information.
Read more
The Art of Not Forgetting A Local Learning Architecture for Continual Learning
Summary
This paper introduces CMP (Cognitive Memory Primitive), a novel continual-learning architecture designed to mitigate catastrophic forgetting by utilizing sparse relational codes and a two-tier competitive memory system. Unlike traditional methods that rely on backpropagation and experience replay, CMP employs local updates and a movement-based plasticity regulation mechanism, allowing it to learn without gradient-derived importance estimates. The authors conducted experiments on a domain-incremental byte-level language modeling protocol, demonstrating that CMP exhibits significantly lower backward transfer compared to a parameter-matched Transformer model trained with online Elastic Weight Consolidation (EWC). The results indicate that CMP's architecture, which integrates sparse representations, local learning, and persistent memory, provides a promising alternative for continual learning. However, the authors also acknowledge a substantial accuracy gap when compared to the Transformer and report negative outcomes in other benchmarks, emphasizing the importance of transparent scientific reporting. Overall, the findings suggest that exploring alternative learning frameworks may enhance continual learning performance.
Methodology
The authors developed CMP by integrating established components such as sparse relational binding, competitive content-addressed memory, and predictive coding principles. They conducted experiments on a controlled domain-incremental protocol, comparing CMP's performance against a Transformer model trained with EWC, while also analyzing the stability of forgetting behavior across different random seeds and domain orders.
Results
CMP demonstrated substantially lower catastrophic forgetting than the Transformer model, with consistent results across multiple experimental conditions. However, it also showed a significant accuracy gap in single-domain tasks and failed to perform well on the Split-MNIST benchmark, indicating limitations in its applicability.
Implications
The findings suggest that architectures designed around local learning and sparse representations could provide new avenues for improving continual learning systems. This approach may complement existing methods and broaden the exploration of learning mechanisms beyond traditional gradient-based optimization.
Prototype Adaptation for Zero-Shot sEMG Movement Classification
Robotics
Time Series
Efficient ML
- Introduction of two novel methods for zero-shot learning in sEMG movement classification: CPI and SAP.
- Geometrical analysis confirms that combined movements can be represented as convex combinations of basic movements in embedding space.
- Creation of a new dataset (BasCom) with a larger variety of movements than existing datasets.
- SAP method shows over 20% accuracy improvement over prior zero-shot learning techniques.
Read more
Prototype Adaptation for Zero-Shot sEMG Movement Classification
Summary
This paper addresses the challenge of movement classification for prosthetic control using surface electromyography (sEMG), particularly focusing on the limitations of current methods that require extensive training data for each new movement. The authors propose two innovative approaches: Compositional Prototype Interpolation (CPI) and Synthetic Adaptation for Prototypes (SAP), which facilitate zero-shot learning of combined movements based solely on training data from basic movements. The methods leverage a linear interpolation assumption in the embedding space of prototype networks, allowing for the creation of new prototypes for unseen movements. The authors provide a geometrical analysis of sEMG signals and demonstrate that combined movements can be approximated as convex combinations of basic movements in the embedding space. They also introduce a novel dataset, BasCom, which includes 11 basic and 8 combined movements, and evaluate their methods on multiple datasets, including NearLab and NinaPro DB3. The results indicate significant improvements in accuracy for combined movements, with SAP outperforming previous zero-shot learning methods by over 20%. This research highlights the potential for reducing the calibration burden in clinical applications of prosthetic control.
Methodology
The authors utilize a prototype network framework to represent combined movements through interpolated prototypes in an embedding space. They adapt these prototypes using synthetic data to maintain robust decision boundaries despite the non-linearities present in sEMG signals. The methods are evaluated using a combination of existing datasets and a newly created dataset, BasCom.
Results
The proposed SAP method significantly outperformed previous zero-shot learning approaches, achieving over 20% higher accuracy in classifying combined movements. The methods were validated across multiple datasets, including those from able-bodied participants and amputees, and demonstrated effectiveness in real-time user studies.
Implications
The findings suggest that the proposed methods can greatly reduce the calibration burden associated with prosthetic control, enabling more efficient and user-friendly applications in clinical settings. This could lead to improved functionality and independence for upper-limb amputees.
DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories
Time Series
- DRIFT integrates direct and recursive forecasting to improve accuracy in ICU physiological predictions.
- The model effectively incorporates treatment interventions, addressing a common limitation in traditional forecasting methods.
- Empirical evaluations show DRIFT outperforms existing models in MAP forecasting across multiple time horizons.
- The framework demonstrates robustness to changes in treatment sequences, maintaining accuracy under altered conditions.
Read more
DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories
Summary
The paper introduces DRIFT, a novel hybrid framework designed for forecasting physiological trajectories in intensive care units (ICUs), specifically focusing on the influence of medical interventions like vasopressors on future vital signs. Traditional forecasting models often fail to effectively incorporate these interventions, leading to inaccuracies. DRIFT combines a direct forecasting model with a recursive, action-conditioned model that provides constrained corrections to improve accuracy. The framework utilizes a Temporal Fusion Transformer (TFT) to generate forecasts based on patient history and treatment sequences, while a shared recursive transition model refines these forecasts through a bounded correction mechanism. Evaluations on two large datasets, MIMIC-IV and eICU-CRD, demonstrate that DRIFT significantly reduces mean absolute error in forecasting mean arterial pressure (MAP) compared to existing models. The results indicate that DRIFT is particularly effective when treatment sequences are altered, showcasing its robustness and adaptability in clinical settings.
Methodology
DRIFT employs a hybrid architecture combining a direct forecasting model using a Temporal Fusion Transformer (TFT) with a recursive, action-conditioned model. The direct model generates forecasts for multiple future time points, while the recursive model provides corrections based on the same action sequence. A stop-gradient distillation technique aligns the outputs of both models, ensuring that the recursive corrections are bounded and do not lead to error accumulation.
Results
Across evaluations on MIMIC-IV and eICU-CRD, DRIFT achieved a mean absolute error reduction of 0.673% for MAP compared to the action-conditioned Temporal Fusion Transformer (TFT-action). The model showed improved accuracy particularly when treatment sequences were altered, with lower MAP errors at 8 and 24 hours. Robustness tests confirmed that DRIFT maintained its performance across various checkpoint-selection criteria.
Implications
The findings suggest that DRIFT can enhance clinical decision-making in ICUs by providing more accurate forecasts of patient trajectories, potentially leading to better management of interventions and improved patient outcomes. The model's ability to adapt to changes in treatment sequences also indicates its practical applicability in dynamic clinical environments.
AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control
Graph Learning
- AgentGFM allows nodes to autonomously control information flow, enhancing adaptability to diverse graph structures.
- The model employs a shared end-to-end trainable policy for all nodes, promoting cross-graph transferability.
- The predict-act-observe-correct process enables nodes to refine their propagation decisions based on feedback.
- Extensive experiments validate the effectiveness of AgentGFM across various graph topologies and transfer scenarios.
Read more
AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control
Summary
The paper introduces AgentGFM, a novel Graph Foundation Model (GFM) that addresses the limitations of existing models in transferring manually designed propagation schemes across diverse graph structures. Traditional GFMs often apply fixed propagation methods that do not adapt to the varying local structural patterns of nodes within graphs. This study proposes a new framework where each node is treated as an agent capable of autonomously determining how information should be propagated through the graph. The proposed model employs a predict-act-observe-correct process, allowing nodes to make decisions regarding source reception, signal-channel selection, and gain-aware halting based on their current state and feedback from the graph. This approach enhances the adaptability of information flow control across different graph topologies. The authors conduct extensive experiments demonstrating the effectiveness of AgentGFM in node-level, graph-level, and large-scale transfer scenarios, showcasing its ability to generalize across diverse graph structures.
Methodology
AgentGFM utilizes an agent-based decision-making framework where each node acts as an agent. The model operates through a predict-act-observe-correct cycle, allowing nodes to predict expected contextual information, make decisions on information reception and propagation, and adjust their states based on observations and discrepancies from predictions. This method enables adaptive information flow control tailored to individual node structures while maintaining a shared policy across nodes.
Results
The experiments conducted demonstrate that AgentGFM significantly outperforms existing GFMs in various tasks, including node-level and graph-level scenarios, as well as in large-scale transfer situations. The model's ability to adaptively control information flow leads to improved performance in diverse graph topologies, confirming its effectiveness in generalizing across different structural patterns.
Implications
AgentGFM has potential applications in various domains involving graph-structured data, such as social network analysis, recommendation systems, and knowledge graph reasoning. Its adaptive information flow control could enhance the performance of graph learning models in real-world applications where graph structures are heterogeneous and dynamic.
Do You Really Need to Pretrain Q-Functions for Online RL Fine-Tuning?
Reinforcement Learning
Robotics
- Naive Q-function pretraining often provides little benefit for online RL fine-tuning.
- The Q-function learned during pretraining does not align with the Q-function learned during online fine-tuning.
- Initialization via Policy Ensemble (IPE) improves Q-function learning by leveraging diverse policy rollouts.
- IPE results in a significant performance boost in fine-tuning tasks compared to naive pretraining.
Read more
Do You Really Need to Pretrain Q-Functions for Online RL Fine-Tuning?
Summary
This paper investigates the necessity of pretraining Q-functions in the context of online reinforcement learning (RL) fine-tuning, particularly when starting from a pretrained policy. The authors challenge the conventional wisdom that suggests pretraining the Q-function on offline data is beneficial. Through systematic empirical analysis, they find that naive Q-function pretraining often yields minimal performance improvements compared to random initialization. The authors attribute this to a fundamental mismatch between the Q-function learned during pretraining and the Q-function that online fine-tuning converges to. To address this issue, they propose a novel method called Initialization via Policy Ensemble (IPE), which trains multiple diverse policies and uses their pooled rollouts to enhance Q-function learning during online RL. The results demonstrate that IPE significantly improves fine-tuning performance, achieving an average of 26% improvement over naive pretraining across various continuous control benchmarks.
Methodology
The authors conducted a systematic empirical analysis to evaluate the impact of Q-function pretraining on online RL fine-tuning. They compared the performance of policies fine-tuned from both pretrained and randomly initialized Q-functions. They introduced IPE, which involves training multiple diverse policies and using their rollouts to bootstrap Q-function learning, thereby increasing the diversity of the data used for training the Q-function.
Results
The study found that naive Q-function pretraining did not lead to significant performance improvements in online RL fine-tuning. In contrast, the proposed IPE method resulted in an average performance improvement of 26% across a suite of challenging continuous control benchmarks compared to naive pretraining.
Implications
The findings suggest that the traditional approach of pretraining Q-functions may not be necessary and that alternative methods like IPE can lead to better performance in online RL fine-tuning. This could influence future research and practices in reinforcement learning, particularly in applications involving pretrained policies.
Data-Dependent Regret and Polyak Corrections for Constrained Online Convex Optimization
Optimization
Theory
- Introduces a refined regret analysis for constrained OCO using data-dependent metrics.
- Identifies a Polyak correction term that enhances the regret bound.
- Proposes AdaOGD-PFS, an adaptive-step-size algorithm achieving improved regret bounds.
- Demonstrates significant empirical improvements in bound performance through experiments.
Read more
Data-Dependent Regret and Polyak Corrections for Constrained Online Convex Optimization
Summary
This paper addresses the problem of constrained online convex optimization (OCO), where a learner aims to minimize regret against adversarially chosen convex costs while adhering to convex constraints. The author proposes an enhancement to the online gradient descent (OGD) algorithm by incorporating a Polyak feasibility step, which allows for efficient projection onto the feasible set using only one constraint evaluation and one subgradient per round. The existing analysis of this approach achieves O(√T) regret but can be improved by adopting a data-dependent perspective. The paper introduces two key refinements: first, it replaces the worst-case gradient envelope with an observed accumulation of gradients, leading to a tighter bound; second, it identifies a Polyak correction term that accounts for the cumulative squared displacement of the feasibility projection, which has not been previously quantified in OCO regret bounds. The proposed AdaOGD-PFS algorithm, which utilizes an adaptive step size, achieves a regret bound of O(√GT), significantly improving upon the traditional O(Gf√T) bound while ensuring per-round constraint satisfaction. Experimental results demonstrate a 38-43% improvement in bound performance, highlighting the contributions of both data-dependent gradients and Polyak corrections.
Methodology
The paper employs a theoretical analysis of the OGD algorithm augmented with Polyak feasibility steps. It replaces standard worst-case assumptions with data-dependent quantities, leading to a refined regret bound. The analysis leverages the strong form of the Pythagorean projection inequality to retain the Polyak correction term, which captures the cumulative effect of feasibility projections.
Results
The refined regret bound established in the paper is significantly tighter than previous bounds, achieving O(√GT) regret compared to O(Gf√T). The empirical results indicate a 38-43% improvement in bound performance across various constrained optimization instances, validating the effectiveness of the proposed methods.
Implications
The findings have important implications for safety-critical applications in fields such as power systems, autonomous control, and clinical decision-making, where maintaining feasibility while minimizing regret is crucial. The proposed methods can enhance decision-making processes in these domains by providing more reliable and efficient optimization strategies.
Sky sphere representation in language models
NLP
Large Language Models
Interpretability
- Most large language models have a decodable representation of the night sky, evidenced by high R2 scores and low angular errors.
- The representation is identified as a spherical feature manifold, marking a significant advancement in the study of high-dimensional representations in LLMs.
- The study employs a systematic approach using prompts focused on celestial proximity, avoiding explicit coordinate references.
- The findings challenge previous notions of geometric representations in LLMs, which have typically been flat or one-dimensional.
Read more
Sky sphere representation in language models
Summary
This paper investigates the representation of the night sky within large language models (LLMs) of approximately 100 billion parameters. The authors analyze whether these models can decode a spherical representation of celestial objects from their residual streams. The study finds that most open-source models exhibit a significant representation of the night sky, often surfacing in the top principal components when prompted with questions about proximity to celestial objects. The representation demonstrates a high degree of accuracy, with leave-one-out (LOO) validation yielding R2 scores between 65% and 85% and median angular errors ranging from 12° to 21°. The authors argue that this representation is not merely a reflection of training data but rather a novel high-dimensional feature manifold. This work contributes to the understanding of how LLMs encode complex spatial relationships and offers insights into the interpretability of their internal representations.
Methodology
The authors utilized a pipeline where various prompts about night sky objects were fed into LLMs in auto-complete mode. They extracted the top 128 principal components of the residual stream vectors from each layer and performed linear regression to fit these components to the true coordinates of celestial objects on a unit sphere. Leave-one-out validation was employed to assess the accuracy of predictions for unseen objects.
Results
The analysis revealed that the positions of celestial objects could be linearly decoded from the residual stream in most models, with R2 scores indicating a variance explanation of 65-85% and median angular errors between 12° and 21°. The representation was prominent in the top principal components, particularly in layers of the models that were analyzed.
Implications
This research has implications for the interpretability of language models, suggesting that they can encode complex spatial relationships in a manner that is both human-inspectable and mathematically rigorous. It opens avenues for further exploration of high-dimensional feature manifolds in AI, potentially enhancing our understanding of how models represent and process spatial information.
SPARC Segmentation to Prediction via Affine Regression and Counterfactuals
Optimization
Interpretability
- Introduces a new synthetic data generation method using Diverse Counterfactual Explanations (DiCE) for better minority-class representation.
- Adapts the PyPARC framework to provide calibrated propensity probabilities for customer segmentation.
- Demonstrates significant improvements in precision and performance over traditional SMOTE-based methods.
- Addresses the unique challenges of B2B transaction propensity prediction, focusing on organizational buying behaviors.
Read more
SPARC Segmentation to Prediction via Affine Regression and Counterfactuals
Summary
This paper addresses the challenges of transaction propensity prediction in B2B e-commerce, where traditional methods like SMOTE fail due to the heterogeneous procurement behaviors of organizations. The authors propose a novel framework that replaces SMOTE with a synthetic data generation approach using Diverse Counterfactual Explanations (DiCE), which produces minority-class samples with better distributional fidelity. Additionally, they adapt the PyPARC piecewise-affine classification framework to generate calibrated propensity probabilities, allowing for effective segmentation of customers into actionable risk tiers. Evaluated on two years of longitudinal data from a large-scale B2B e-commerce platform, the proposed architecture achieves a precision of 93.1% at a decision threshold of 0.8, significantly outperforming SMOTE-based baselines. This framework enhances the ability of businesses to conduct high-precision marketing campaigns, improving customer activation and return on investment.
Methodology
The authors developed a framework that integrates a DiCE-based augmentation strategy to generate synthetic minority-class samples while preserving feasible B2B account states. They also modified the PyPARC classification framework to output calibrated propensity probabilities instead of discrete class predictions, facilitating better customer segmentation.
Results
The proposed framework achieved 93.1% precision at a decision threshold of 0.8, a 9.2-percentage-point improvement over SMOTE-based baselines at the same threshold and a 26.1-point improvement at threshold 0.7. These results indicate consistent superiority across various operating points.
Implications
The findings suggest that the proposed framework can significantly enhance the effectiveness of marketing campaigns in B2B e-commerce by enabling more accurate predictions of customer behavior, leading to improved resource allocation and customer engagement strategies.
Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation
Large Language Models
NLP
Theory
- Introduces the concept of representational lossiness in CATE estimation.
- Develops CURL, an uncertainty-guided representation learning adapter for CATE.
- Demonstrates improved performance in heterogeneous treatment effect estimation across multiple benchmarks.
- Utilizes large language models to enhance the representation of covariate information.
Read more
Uncertainty-Guided LLM Semantic Augmentation for Heterogeneous Treatment Effect Estimation
Summary
This paper addresses the challenge of estimating heterogeneous treatment effects (HTE), particularly the conditional average treatment effect (CATE), which is crucial for personalized interventions in fields like precision medicine and targeted marketing. The authors identify a key issue known as 'representational lossiness', where the raw encoding of covariates fails to capture essential semantic relationships and higher-order interactions, leading to instability in local estimates. To mitigate this, they propose CURL (Causal Uncertainty-guided Representation Learning), a plug-in adapter that leverages estimator uncertainty to guide the allocation of pretrained semantic capacity from large language models (LLMs) to improve representation learning. CURL employs two role-conditioned prompts to create distinct representations for assignment and heterogeneity, which are then routed through separate pathways in the estimation process. The effectiveness of CURL is demonstrated across four benchmarks, where it enhances the performance of ten different host learners, supported by various analyses that validate its design and functionality.
Methodology
The authors developed CURL, which uses uncertainty from propensity and CATE estimates to prioritize samples for semantic augmentation. It constructs separate representations for assignment and heterogeneity from observed covariates and routes them through distinct pathways in the estimator. The method involves querying a frozen LLM with role-conditioned prompts and updating the augmented set as the host estimator evolves.
Results
CURL significantly improves the performance of ten host learners across four benchmark datasets, demonstrating its effectiveness in addressing representational lossiness. The results indicate that the proposed method successfully enhances the stability and accuracy of CATE estimates by leveraging semantic knowledge from LLMs.
Implications
The findings suggest that integrating LLMs into causal inference frameworks can enhance the estimation of treatment effects, potentially leading to more effective personalized interventions in various domains such as healthcare and marketing. The approach may also inspire further research into the use of pretrained models for structured data tasks.
Existence-Field Diffusion Model for Spatial Point Processes with Variable Cardinality
Generative Models
- Introduces the Existence-Field Diffusion Model (EFDM) for spatial point processes with variable cardinality.
- EFDM uses existence variables to represent the degree of presence of potential points, allowing for a unified diffusion process.
- The model avoids discrete transitions, providing a more flexible approach to generative modeling.
- Demonstrates improved performance on datasets with varying cardinality compared to existing methods.
Read more
Existence-Field Diffusion Model for Spatial Point Processes with Variable Cardinality
Summary
This paper presents the Existence-Field Diffusion Model (EFDM) for generative modeling of spatial point processes (SPP) characterized by variable cardinality and spatial configurations. Traditional methods either decouple the modeling of cardinality and spatial structure or rely on discrete operations to modify the number of points, leading to inflexible generative dynamics. The proposed EFDM introduces an existence variable for each potential point, representing its degree of presence, allowing for a unified diffusion process that simultaneously models both spatial locations and cardinality. This continuous approach avoids the need for explicit discrete transitions, enabling a more flexible and general framework for SPP modeling. The authors demonstrate the effectiveness of EFDM on synthetic and real-world datasets, showing that it captures the dependency between cardinality and spatial structure, outperforming existing methods in conditional evaluations across varying cardinality levels.
Methodology
The EFDM employs a continuous diffusion process where both the positions of points and their existence degrees are diffused to noise in the forward process and jointly denoised in the reverse process. This allows for probabilistic sampling of existence at the end of generation, enabling a natural adjustment of cardinality during the diffusion.
Results
The EFDM was evaluated on both synthetic and real-world datasets, showing significant improvements in modeling capabilities, particularly in capturing the relationship between cardinality and spatial structure. The model achieved stronger performance in conditional evaluations across different cardinality levels compared to previous approaches.
Implications
The EFDM has potential applications in various fields where spatial point processes are relevant, such as urban planning, environmental modeling, and molecular structure analysis. Its ability to model variable cardinality and spatial configurations could enhance generative modeling tasks in these domains.
From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
Generative Models
- Introduces a diffusion-based generative model for air quality reconstruction.
- Demonstrates joint modeling of multiple pollutants to capture inter-pollutant correlations.
- Achieves high accuracy in reconstructing pollution fields from sparse observations.
- Utilizes data augmentation techniques for improved generalization to real-world data.
Read more
From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
Summary
This paper addresses the challenge of reconstructing urban air quality from sparse observations of pollutants such as NO2, O3, PM2.5, and PM10. Traditional deterministic models face limitations due to uncertainties in emission inventories and meteorological data. The authors propose a novel diffusion-based generative framework that enhances the reconstruction of air pollution fields by leveraging deep learning techniques. The model is trained on simulated data and evaluated against real-world observations from monitoring stations in Paris. The study highlights the importance of joint modeling of multiple pollutants and introduces data augmentation methods to improve generalization to real-world scenarios. The results indicate that the generative model achieves high structural similarity and realistic spatial patterns, outperforming deterministic models. This work represents a significant advancement in the application of generative AI for environmental monitoring, demonstrating the potential for real-time deployment in air quality reconstruction tasks.
Methodology
The authors employed a diffusion-based generative model to reconstruct air quality fields from sparse sensor observations. They trained the model on simulated full-field data and evaluated its performance using real-world observations from Paris. Voronoi tessellations were used to create a spatially consistent representation of the sensor network, addressing issues of temporally missing data. The evaluation included point-wise cross-validation and spectral analysis to assess model performance.
Results
The diffusion-based generative model demonstrated high structural similarity with simulated validation data and produced realistic spatial patterns when applied to real-world observations. The model effectively handled noisy data and strong spatial variability, achieving better performance compared to traditional deterministic models. The findings suggest that the model generalizes well to unseen data, indicating its robustness for practical applications.
Implications
This research has significant implications for urban air quality monitoring and public health decision-making. The ability to accurately reconstruct pollution fields from sparse observations can enhance exposure assessment and inform policy measures. The generative approach also opens avenues for real-time monitoring and response to air quality issues, contributing to improved environmental health outcomes.
Dynamic Parameterization Is Not Dynamic Inference
Theory
Efficient ML
Large Language Models
- Dynamic parameterization does not equate to dynamic inference.
- Frozen-Controller Auditing (FCA) effectively isolates the effects of coefficient assignment on model performance.
- Static layerwise profiles can closely approximate model performance despite dynamic coefficients.
- Functional dependence on coefficient assignment is critical for understanding model behavior.
Read more
Dynamic Parameterization Is Not Dynamic Inference
Summary
This paper addresses the misconception that input-dependent controller coefficients in machine learning models imply dynamic inference or computational efficiency. The authors clarify three distinct properties: coefficient variation, the dependence of a frozen model on coefficient assignment, and conditional execution. They introduce a method called Frozen-Controller Auditing (FCA), which involves caching the coefficient tensor during an unperturbed trajectory, disabling the controller, and replaying the model with various coefficient assignments. The study evaluates two model families, FeatureGate and MUDDPythia, demonstrating that while dynamic parameterization exists, it does not necessarily lead to dynamic inference or computational savings. The findings emphasize the need for separate assessments of coefficient variation, functional dependence, and execution in claims about dynamic models.
Methodology
The authors propose the Frozen-Controller Auditing (FCA) method, which involves caching the coefficient tensor along an unperturbed trajectory, disabling the controller, and replaying the model with cross-input reassignment, token shuffling, and static profiles derived from an independent calibration set. This approach allows for the evaluation of the functional dependence of the model without feedback from recomputed controllers.
Results
The application of FCA to FeatureGate models showed that static layerwise profiles retained 98.70% and 99.43% of the Correct-to-GlobalMean performance gap across different model sizes. In contrast, MUDDPythia demonstrated a strong dependence on the alignment of cross-layer coefficients with their original inputs, resulting in significant performance penalties when coefficients were reassigned or shuffled. Both models executed every Transformer block, indicating that dynamic coefficients alone do not lead to computational savings.
Implications
The findings suggest that researchers and practitioners should be cautious when interpreting dynamic models. The distinction between coefficient variation and actual dynamic inference is crucial for understanding model efficiency and performance. This work encourages a more rigorous evaluation of dynamic models in machine learning, potentially influencing future model design and assessment strategies.
FloDR: An invertible dimensionality reduction method based on a normalising flow
Theory
Interpretability
Generative Models
- FloDR preserves both local and global structures in high-dimensional data embeddings.
- The method retains additional coordinates for diagnostic purposes, enhancing interpretability.
- Two diagnostic fields, conditional spread and hidden contrast, provide insights into the embedding quality.
- FloDR's performance is competitive with existing methods like t-SNE and UMAP.
Read more
FloDR: An invertible dimensionality reduction method based on a normalising flow
Summary
The paper introduces FloDR, a novel dimensionality reduction technique that utilizes an invertible normalising flow to create two-dimensional embeddings of high-dimensional data. Unlike traditional methods such as t-SNE and UMAP, which often discard information during the optimization process, FloDR retains additional coordinates that provide valuable insights into the data structure. The method employs a bijection built from affine coupling layers, allowing for both local and global structure preservation in the embeddings. Additionally, FloDR generates diagnostic visualizations, including the conditional spread and hidden contrast fields, which quantify the uncertainty and information loss associated with the embeddings. These fields are empirically validated through a certification process, ensuring their reliability. The results demonstrate that FloDR competes well with established methods while offering enhanced interpretability and diagnostic capabilities.
Methodology
FloDR employs an invertible normalising flow architecture, specifically using affine coupling layers to create a bijection from high-dimensional data to a two-dimensional embedding. The first two output coordinates are used for the embedding, while the remaining coordinates are retained for evaluation. The method incorporates neighbourhood and ordinal losses during training to optimize the layout. Additionally, it generates diagnostic fields based on the conditional distribution of the data given the embedding position, validated through a held-out certification process.
Results
FloDR demonstrates competitive performance against established dimensionality reduction techniques like t-SNE and UMAP. The method effectively retains information that is typically lost in other approaches, allowing for more accurate representations of both local and global data structures. The diagnostic fields produced by FloDR provide measurable insights into the embedding's reliability, with empirical tests confirming their accuracy.
Implications
FloDR's ability to retain and visualize additional data dimensions enhances the interpretability of high-dimensional data embeddings, making it a valuable tool for researchers in fields such as bioinformatics, data visualization, and machine learning. Its diagnostic capabilities can help users make more informed decisions based on the embeddings, potentially leading to better insights in various applications.
Lantern: Conflict-Aware Gradient Blending for Physics-Guided Diffusion Models in Calorimeter Simulation
Generative Models
Optimization
Theory
- Introduction of the Correlation Frobenius Distance (CFD) for evaluating correlation fidelity in calorimeter simulations.
- Development of two physics-aware auxiliary losses to enhance the training of diffusion models.
- Implementation of GradBlend, a novel gradient blending technique that prioritizes denoising while incorporating physics guidance.
- Significant improvements in simulation fidelity metrics (FPD and CFD) compared to traditional denoising methods.
Read more
Lantern: Conflict-Aware Gradient Blending for Physics-Guided Diffusion Models in Calorimeter Simulation
Summary
This paper addresses the computational challenges in simulating calorimeter showers, which are critical for high-energy physics experiments, particularly at the High-Luminosity LHC. Traditional Monte Carlo methods, while accurate, are computationally expensive and impractical for the required data volumes. The authors propose 'Lantern', a physics-guided diffusion model that incorporates physics-aware auxiliary losses to improve the fidelity of simulated showers. They introduce the Correlation Frobenius Distance (CFD) as a novel metric to evaluate the correlation fidelity of simulated showers at both layer-wise and voxel-wise scales. The methodology involves two auxiliary losses: a variance-stabilized voxel residual loss and a graph Laplacian loss, which are integrated into the training process using a new gradient blending technique called GradBlend. This approach allows the model to maintain the generative objective while incorporating physics constraints. The results demonstrate that Lantern significantly outperforms traditional denoising methods, achieving improvements in both the Fréchet Physics Distance (FPD) and the newly introduced CFD metric. The findings suggest that incorporating physics-aware structures into generative models can enhance their performance in complex simulations.
Methodology
The authors developed a physics-guided diffusion model named Lantern, which integrates two auxiliary losses into the training process. The first loss is a variance-stabilized voxel residual loss that captures the statistical properties of voxel energies, while the second is a graph Laplacian loss that accounts for the spatial relationships between neighboring voxels. These losses are combined using GradBlend, a gradient blending rule that anchors the update magnitude to the denoising gradient while allowing the auxiliary losses to influence the update direction.
Results
The experiments conducted on Dataset 2 of the CaloChallenge showed that Lantern, using GradBlend and the graph Laplacian loss, improved FPD and CFD metrics significantly compared to traditional denoising methods. The auxiliary losses were shown to enhance the model's ability to reproduce the physical characteristics of calorimeter showers, with improvements in fidelity ranging from 2 to 100 times relative to denoising alone.
Implications
The findings suggest that integrating physics-aware structures into generative models can lead to more accurate simulations in high-energy physics, potentially improving detector design, calibration, and event reconstruction processes. This approach could be applied to other domains where physical fidelity is crucial in generative modeling.
RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment
Time Series
Efficient ML
Large Language Models
- RAG-HAR+ reframes LLM usage in HAR as a two-stage process, enhancing efficiency.
- The offline Retrieval Designer Agent customizes feature groups for improved retrieval.
- Majority voting is used for confident predictions, reducing LLM dependency.
- The framework maintains competitive performance while lowering computational costs.
Read more
RAG-HAR+: Towards Cost-Efficient LLM-Based Human Activity Recognition for Edge Deployment
Summary
The paper introduces RAG-HAR+, an innovative framework for Human Activity Recognition (HAR) that leverages Large Language Models (LLMs) in a cost-efficient manner suitable for edge deployment. Traditional deep learning methods for HAR require extensive dataset-specific training and are not easily adaptable to new sensor settings or activity types. RAG-HAR+ addresses these challenges by employing a retrieval-augmented approach that minimizes reliance on LLM inference. The framework includes an offline Retrieval Designer Agent that creates dataset-specific feature groups from a diverse set of motion descriptors, enhancing the retrieval process. During inference, RAG-HAR+ utilizes majority voting among retrieved neighbors for confident predictions, only resorting to the LLM for ambiguous cases. This two-stage process significantly reduces computational costs and latency, making it more feasible for continuous monitoring on resource-constrained devices. The authors validate RAG-HAR+ across six HAR benchmarks, demonstrating competitive or improved performance while decreasing LLM usage and inference time. Additionally, they extend the RAG-HAR mobile prototype to showcase the practical application of their method in mobile sensing scenarios.
Methodology
RAG-HAR+ employs a two-stage approach where an offline Retrieval Designer Agent creates dataset-specific feature groups from various motion descriptors. During inference, it uses majority voting over retrieved neighbors for classification, invoking an LLM only for ambiguous cases. This method reduces the computational burden and latency associated with continuous LLM usage.
Results
The framework was tested across six HAR benchmarks, showing that RAG-HAR+ maintains or improves performance compared to existing methods while significantly reducing LLM usage, token consumption, and inference time.
Implications
RAG-HAR+ has significant implications for real-time HAR applications in healthcare, fitness tracking, and smart environments, particularly in resource-constrained settings. Its cost-efficient design allows for scalable and adaptable HAR systems without the need for extensive retraining.
Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark
Theory
Optimization
Efficient ML
- Marginal conformal prediction shows severe undercoverage for minority classes, dropping below 1%.
- Mondrian conformal prediction restores minority coverage, achieving an average improvement of 61.7 percentage points.
- Cost-controlled abstention reduces overall expected decision costs by allowing deferral of ambiguous predictions.
- Dataset-specific break-even thresholds for human review costs are established, guiding cost-effective decision-making.
Read more
Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark
Summary
This paper addresses the challenges of high-stakes decision-making in imbalanced datasets, particularly in fields like credit scoring and healthcare, where errors can have severe consequences. It highlights the limitations of standard marginal conformal prediction (CP), which often fails to provide adequate coverage for minority classes, sometimes dropping below 1%. The authors benchmark various conformal prediction methods, including marginal CP and Mondrian CP, across 15 real-world imbalanced datasets. They find that Mondrian CP significantly improves minority class coverage by an average of 61.7 percentage points. Additionally, the paper introduces a cost-controlled abstention mechanism that allows for deferring ambiguous predictions to human experts, thereby reducing expected decision costs. The authors also derive dataset-specific break-even thresholds for human review costs, providing actionable insights for deploying cost-sensitive decision support systems.
Methodology
The authors conducted a comprehensive benchmark comparing marginal conformal prediction, Mondrian conformal prediction, and cost-controlled abstention mechanisms across 15 imbalanced datasets. They utilized 7 classification models, 3 probability calibration techniques, and performed 3,150 experimental runs to evaluate the performance of these methods in terms of minority class coverage and decision costs.
Results
The study found that marginal conformal prediction significantly underperformed in terms of minority class coverage, while Mondrian conformal prediction provided a substantial improvement. The integration of cost-controlled abstention mechanisms led to a reduction in expected decision costs, demonstrating the effectiveness of deferring ambiguous predictions to human experts. The derived break-even thresholds for human review costs offer practical guidelines for decision support systems.
Implications
The findings have significant implications for high-stakes decision-making systems, suggesting that incorporating cost-sensitive conformal prediction and human-in-the-loop mechanisms can enhance decision accuracy and reduce costs. This approach can be particularly beneficial in fields where the consequences of misclassification are severe, such as healthcare and finance.
Mechanisms of Width Scaling in Normalized Residual Networks: The Effective Alignment Dimension
Theory
Optimization
- Introduction of the effective alignment dimension as a measure of signal-noise geometry in activation gradients.
- Derivation of a finite-sample upper bound on misalignment probability based on effective alignment dimension and sample size.
- Integration of the effective alignment dimension into the train-test expansion framework for improved risk assessment.
- Empirical validation showing that wider models exhibit better alignment and lower misalignment.
Read more
Mechanisms of Width Scaling in Normalized Residual Networks: The Effective Alignment Dimension
Summary
This paper addresses the challenge of understanding how width scaling in neural networks, particularly normalized residual networks, affects their performance on unseen data. The authors introduce the concept of the effective alignment dimension, which quantifies the geometry of signal and noise in activation gradients. They derive an exact mean and variance for the inner product of training and test gradients, leading to a finite-sample upper bound on misalignment probability. This bound is dependent on the effective alignment dimension and effective sample size, without requiring strict assumptions about covariance or growth rates. The authors integrate this new certificate into the existing framework for train-test residual expansion, providing a high-probability condition for test-risk improvement. Empirical experiments on various models, including LLaMA-style Transformers and ResNet-20, demonstrate that wider models tend to have larger effective alignment dimensions and lower empirical misalignment. The study concludes that the alignment statistic can predict changes in held-out loss, confirming the practical relevance of their theoretical findings.
Methodology
The authors derive an exact second-moment characterization of the alignment between training and test gradients. They analyze the alignment statistic directly rather than separately controlling training and test deviations. The effective alignment dimension is calculated from the activation gradients, and experiments are conducted on various neural network architectures to validate the theoretical findings.
Results
The study finds that the effective alignment dimension generally increases with model width, leading to reduced empirical misalignment. Direct interventions in residual connections confirm that positive alignment correlates with favorable changes in held-out loss, supporting the theoretical framework established in the paper.
Implications
The findings suggest that understanding the effective alignment dimension can guide the design of neural networks, particularly in determining the benefits of width scaling. This could lead to more effective model growth strategies and improved generalization on unseen data.
Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning
Federated Learning
Interpretability
- Federated Learning is utilized to ensure patient data privacy while predicting CKD.
- Multiple ensemble methods (Random Forest, AdaBoost, XGBoost) are compared to identify the best-performing model.
- Explainable AI techniques, particularly LIME, are integrated to enhance model interpretability.
- The global model achieved an average accuracy of 99%, indicating high predictive performance.
Read more
Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning
Summary
This paper addresses the critical issue of early detection of Chronic Kidney Disease (CKD) through a novel approach that combines Explainable Artificial Intelligence (XAI) with Federated Learning (FL). CKD is a significant public health concern characterized by the gradual loss of kidney function, necessitating timely diagnosis to prevent severe complications. The authors employed a VotingClassifier in a federated learning framework, integrating multiple algorithms including Random Forest, AdaBoost, and XGBoost to identify the most effective model for CKD prediction. The study utilized GridSearchCV for hyperparameter optimization on the client side, ensuring enhanced model performance. A key aspect of the research is the incorporation of XAI techniques, specifically LIME, to provide transparency and interpretability of the model's predictions. The results demonstrated an impressive average accuracy of 99% for the global model, underscoring the potential of interpretable federated learning models in facilitating early CKD diagnosis and promoting data-driven healthcare solutions. This research highlights the importance of maintaining patient data privacy while leveraging advanced machine learning techniques for healthcare applications.
Methodology
The authors implemented a federated learning framework using a VotingClassifier that combined Random Forest, AdaBoost, and XGBoost algorithms. They applied GridSearchCV for hyperparameter tuning on the client side and utilized LIME for model interpretability.
Results
The global model achieved an average accuracy of 99%, demonstrating the effectiveness of the federated learning approach combined with ensemble methods for CKD prediction.
Implications
The findings suggest that federated learning can be a viable approach for developing predictive models in healthcare while ensuring data privacy. The integration of explainable AI techniques can enhance trust in AI-driven healthcare solutions, facilitating better decision-making and patient outcomes.
Calibrated Partial Resets: Preventing Policy Collapse in Continual Reinforcement Learning
Reinforcement Learning
Optimization
Robotics
- Introduces Calibrated Partial Resets (CPR) as a solution to policy collapse in continual RL.
- CPR selectively adjusts low-utility neurons based on their utility, avoiding abrupt changes.
- Demonstrates superior performance over existing methods in multiple RL benchmarks.
- Establishes a tunable trade-off between network plasticity and peak performance.
Read more
Calibrated Partial Resets: Preventing Policy Collapse in Continual Reinforcement Learning
Summary
This paper addresses the challenge of policy collapse in continual reinforcement learning (RL) caused by the accumulation of dormant neurons and loss of expressivity in neural networks during non-stationary training. The authors propose a novel approach called Calibrated Partial Resets (CPR), which periodically adjusts low-utility neurons towards their initialization without fully reinitializing them. This method aims to maintain the plasticity of the network while avoiding the instability associated with binary reset methods. CPR selectively applies a continuous coefficient based on each neuron's utility, allowing for smoother adjustments compared to uniform decay methods. The authors demonstrate that CPR effectively prevents policy collapse over extensive training periods and outperforms existing decay and reset-based methods across various benchmarks, including SlipperyAnt, Continual MetaWorld, and Continual MinAtar. The findings highlight the importance of utility-scaled reinitialization in balancing plasticity and performance in continual learning settings.
Methodology
The authors developed CPR, an optimizer that periodically pulls low-utility neurons towards their initialization using a scaling factor based on each neuron's utility. This method contrasts with binary resets, which can destabilize training, and uniform decay methods, which apply indiscriminate changes. The effectiveness of CPR was evaluated through extensive experiments across various RL environments, measuring performance and the incidence of policy collapse.
Results
CPR was the only method to avoid policy collapse over 400 million training steps in the SlipperyAnt environment. It also outperformed prior decay and reset-based methods in other benchmarks, achieving higher episodic returns and demonstrating a robust performance across multiple seeds.
Implications
The findings suggest that utility-scaled reinitialization can significantly enhance the stability and adaptability of RL agents in continual learning scenarios. This approach may be applicable in various domains where agents must learn and adapt over time, such as robotics and autonomous systems.
Parameter-Free Dynamic Regret for Online Convex Optimization under Heavy-Tailed Noise
Optimization
Theory
- Introduces HT-PAder, a parameter-free algorithm for OCO under heavy-tailed noise.
- Achieves optimal universal dynamic regret without prior knowledge of problem parameters.
- Extends static regret results to dynamic regret in non-stationary environments.
- Establishes a matching lower bound, confirming the optimality of the proposed algorithm.
Read more
Parameter-Free Dynamic Regret for Online Convex Optimization under Heavy-Tailed Noise
Summary
This paper addresses the challenge of achieving parameter-free universal dynamic regret in online convex optimization (OCO) under heavy-tailed noise, where the stochastic gradient oracle has a finite p-th central moment for p in (1, 2]. The author proposes a novel algorithm called HT-PAder, which combines restarted AdaGrad experts with a pathwise meta-algorithm, AdaGrad-Hedge. This approach does not require prior knowledge of problem parameters such as the Lipschitz constant or noise level. The algorithm achieves an expected universal dynamic regret that is optimal in terms of the path-length exponent. The paper also establishes a matching lower bound, confirming the optimality of the proposed method. The results extend previous findings on static regret to the dynamic setting, demonstrating that tracking rapidly changing comparators under heavy-tailed noise is fundamentally more challenging.
Methodology
The methodology involves a combination of a pathwise meta-algorithm (AdaGrad-Hedge) and restarted AdaGrad experts. The pathwise nature of the meta-algorithm allows it to handle heavy-tailed gradient oracles effectively. The restarted AdaGrad experts adapt to unknown path lengths and gradient scales, ensuring parameter-freeness. The algorithm's performance is analyzed through a careful examination of regret bounds and the introduction of a novel adversarial environment for lower bound proofs.
Results
HT-PAder achieves an expected universal dynamic regret of eO(GD^p T(1 + PT/D) + σDT^(1/p)(1 + PT/D)^((p-1)/p)). The results show that the algorithm matches the minimax rate for finite variance (p = 2) and provides a weaker dependence on path length for p < 2, reflecting the challenges posed by heavy-tailed noise. A matching lower bound is also established, confirming the optimality of the proposed exponent.
Implications
The findings have significant implications for online learning in environments characterized by heavy-tailed noise, which is common in various real-world applications such as finance and deep learning. The parameter-free nature of the algorithm makes it particularly useful in practical scenarios where prior knowledge of problem parameters is unavailable.
Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs
Graph Learning
Time Series
- Introduces a novel framework for modeling longitudinal disease trajectories as temporal graphs.
- Utilizes contrastive graph neural networks to learn patient observation representations.
- Implements structure-aware random walks to enhance the quality of learned embeddings.
- Demonstrates improved clustering performance on various longitudinal biomedical datasets.
Read more
Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs
Summary
This paper addresses the challenges of understanding disease trajectories from longitudinal clinical data, which often exhibit complex temporal dynamics and involve heterogeneous patient cohorts. The author proposes a novel contrastive representation learning framework, termed RankWalk, which models multivariate disease trajectories as temporal graphs. In this framework, nodes represent patient observations over time, while edges capture both temporal continuity and structural similarity between trajectories. The approach utilizes structure-aware random walks to guide contrastive learning, generating embeddings that maintain temporal context and trajectory topology. The learned representations facilitate robust clustering of patients with similar disease progression patterns and uncover latent structures in longitudinal data. The proposed method is evaluated through simulation studies and real-world biomedical datasets, demonstrating significant improvements in clustering performance over existing longitudinal clustering techniques.
Methodology
The RankWalk framework constructs a heterogeneous graph where each observation is a node. Temporal edges connect consecutive observations of the same subject, while similarity edges link subjects with comparable characteristics. A weighted contrastive objective is employed to learn node representations, emphasizing informative positive pairs. The framework is evaluated using simulation studies and real-world datasets to assess its clustering capabilities.
Results
The experimental results indicate that the RankWalk framework consistently outperforms traditional longitudinal clustering methods, showcasing enhanced clustering performance across both simulated and real-world biomedical datasets.
Implications
This work has potential applications in patient stratification, disease subtyping, and improving the understanding of disease progression patterns in clinical settings. The framework can be adapted for various domains where longitudinal data analysis is crucial.
Rethinking CD: A Reproducibility Study and Extension on the Ineffectiveness of Contrastive Decoding at Mitigating Object Hallucinations in MLLMs
Multimodal
Large Language Models
Computer Vision
- Contrastive Decoding (CD) does not effectively mitigate object hallucinations in MLLMs.
- The performance gains from CD are often spurious and can be replicated by non-visual controls.
- The adaptive plausibility constraint (APC) leads to a degradation of sampling quality.
- CD's adjustments fail to selectively suppress hallucinated tokens, often amplifying them instead.
Read more
Rethinking CD: A Reproducibility Study and Extension on the Ineffectiveness of Contrastive Decoding at Mitigating Object Hallucinations in MLLMs
Summary
This paper conducts a reproducibility study and extension of previous work questioning the effectiveness of Contrastive Decoding (CD) in mitigating object hallucinations in multimodal large language models (MLLMs). The authors reproduce and extend the findings of Yin et al. (2026), which suggested that the performance gains attributed to CD may not reflect genuine improvements in visual grounding. The study confirms that CD induces a unidirectional output distribution shift and that the adaptive plausibility constraint (APC) reduces sampling to greedy search. Through a series of experiments, the authors analyze the logit distributions from various CD strategies, propose a proxy method for comparison, and investigate the propagation of hallucination signals through model layers. The results indicate that the improvements from CD are often spurious, failing to consistently enhance visual grounding and reduce hallucinations. The findings challenge the current reliance on CD strategies and highlight the need for more effective methods to address hallucinations in MLLMs.
Methodology
The authors reproduced the findings of Yin et al. (2026) using newer models and datasets. They conducted experiments to analyze logit distributions, compare CD with a proxy method, and perform layer-wise examinations of hallucination propagation. Metrics such as flip precision and hallucination-injection rates were used to assess the effectiveness of CD.
Results
The study confirmed that CD induces a bias towards 'Yes' outputs, with accuracy gains matched or exceeded by non-visual controls. The APC was shown to degrade sampling quality, and the analysis revealed that CD often amplifies hallucinations rather than mitigating them. Overall, the apparent improvements from CD were found to be statistically indistinguishable from those achieved by other methods.
Implications
These findings suggest that current contrastive decoding strategies may not be reliable for reducing hallucinations in MLLMs, indicating a need for the development of more effective techniques. This has implications for applications in high-stakes environments where accurate visual grounding is critical, such as autonomous driving and healthcare.
Stable FP4 Training via Transposition-Invariant Block Quantization
NLP
Large Language Models
Efficient ML
- Identified transposition-induced scale inconsistency as a major source of instability in FP4 training.
- Proposed a 2D block FP4 quantization framework that ensures consistent scaling across forward and backward passes.
- Combined FP4 linear layers with MXFP8 attention for a practical mixed-precision training approach.
- Achieved stable training for models up to 30 billion parameters with minimal performance degradation.
Read more
Stable FP4 Training via Transposition-Invariant Block Quantization
Summary
This paper addresses the challenges of training large language models (LLMs) using 4-bit floating point (FP4) precision, which has been hindered by instability during optimization. The authors identify a critical issue in existing microscaling approaches, where tensor transposition leads to scale inconsistency between forward and backward passes, resulting in biased gradient updates. To resolve this, they propose a novel framework utilizing 2D block FP4 quantization that maintains transposition-invariant scaling, ensuring consistent behavior across computations. Additionally, they incorporate truncation-free scaling and stochastic rounding to minimize quantization errors and preserve unbiased gradients. The method is further enhanced by employing MXFP8 quantization for attention mechanisms, allowing for a mixed-precision approach. The proposed framework is evaluated on dense LLMs with up to 7 billion parameters and a 30 billion parameter Mixture-of-Experts model, demonstrating stable FP4 training with performance closely matching BF16 precision, showing less than 1.3% degradation in perplexity and downstream accuracy. This work provides a pathway to more efficient LLM training by enabling practical FP4 training at scale.
Methodology
The authors developed a transposition-invariant FP4 training framework based on 2D block quantization. This approach ensures that the same scaling factors are used for both forward and backward passes, eliminating scale inconsistency. They also employed truncation-free scaling and stochastic rounding to manage quantization errors. For attention mechanisms, MXFP8 quantization was utilized to stabilize the training process.
Results
The proposed method achieved stable end-to-end FP4 training across various model sizes, including dense LLMs up to 7 billion parameters and a 30 billion parameter Mixture-of-Experts model. The performance closely matched that of BF16 precision, with less than 1.3% degradation in perplexity and downstream accuracy, demonstrating the effectiveness of the approach.
Implications
This research has significant implications for the efficient training of large language models, as it provides a viable pathway to utilize lower precision formats like FP4 without sacrificing stability or performance. This could lead to reduced computational costs and energy consumption in training large-scale models.
A Machine-Learning-Based Gas Lift Optimization Workflow for Unconventional Fields
Optimization
Time Series
- Development of a machine learning model to forecast Gas Lift Performance Curves without the need for downhole data.
- Integration of Bayesian Optimization to determine optimal gas injection rates under facility constraints.
- Pilot study showed over 5% production uplift across 30 wells, leading to full deployment in over 200 wells.
- The workflow is designed to be cost-effective and applicable in unconventional fields with limited data availability.
Read more
A Machine-Learning-Based Gas Lift Optimization Workflow for Unconventional Fields
Summary
This paper introduces an automated, data-driven workflow utilizing machine learning (ML) for optimizing gas lift operations in unconventional oil fields, specifically targeting the Bakken formation. The proposed workflow integrates a machine learning model that predicts the Gas Lift Performance Curve based on historical production data, eliminating the need for downhole gauges or multi-rate well tests. The model is coupled with a Bayesian Optimization Framework to determine the optimal gas injection rates while adhering to facility constraints. The pilot implementation across 30 wells demonstrated an average production uplift of over 5%. Following this success, the workflow has been fully deployed across more than 200 gas lift and plunger-assisted gas lift wells in Bakken. This approach offers a cost-effective and efficient solution for gas lift optimization in environments where traditional data collection methods are impractical due to cost or operational constraints.
Methodology
The methodology consists of three main components: (1) a machine learning forecaster trained on historical production and gas injection data to predict Gas Lift Performance Curves, (2) a Bayesian Optimization framework to compute optimal gas injection rates, and (3) a deployment strategy that applies the model to real-time operations without requiring downhole measurements.
Results
The implementation of the ML-based gas lift optimization workflow resulted in an average production uplift of over 5% during the pilot phase across 30 wells. Following this, the workflow was successfully deployed across more than 200 wells, demonstrating its scalability and effectiveness in optimizing gas lift operations.
Implications
The findings suggest that the proposed workflow can significantly enhance production efficiency in unconventional oil fields, particularly in scenarios where traditional data collection methods are not feasible. This approach could be adapted for use in other unconventional assets, potentially leading to broader applications in the oil and gas industry.
Learned, Relied Upon, or Necessary? Separating Checkpoint Dependence from Task-Level Value in Sheaf GNNs
Graph Learning
Theory
Interpretability
- Introduces a framework to differentiate between checkpoint reliance and task-level value in sheaf GNNs.
- Demonstrates that learned maps can influence model performance without being essential for task success.
- Presents a task-null theorem explaining the divergence between reliance and task value.
- Conducts experiments across multiple benchmarks to validate the proposed framework.
Read more
Learned, Relied Upon, or Necessary? Separating Checkpoint Dependence from Task-Level Value in Sheaf GNNs
Summary
This paper addresses the interpretation of learned restriction maps in sheaf graph neural networks (GNNs), challenging the assumption that their presence indicates useful edge geometry. The author distinguishes between checkpoint reliance and task-level value through two estimands: checkpoint reliance, which evaluates the impact of learned maps on a fixed predictor, and protocol-relative replacement, which assesses the ability of retrained models to recover task performance after removing map capacity. The study introduces a task-null theorem that reveals the divergence between these claims, highlighting that labels only identify certain classifier directions while leaving other degrees of freedom unaccounted for. The paper presents an exact frame model to delineate when reliance on learned transport becomes unnecessary. Through experiments on various benchmarks, including NSD and DNSD implementations, the author demonstrates that while learned maps can influence performance, they are not always essential for task success. The findings suggest that claims regarding learned transport should be validated through both checkpoint interventions and matched retraining.
Methodology
The author employs a theoretical framework to separate the effects of learned maps into three distinct claims: map movement, checkpoint reliance, and protocol-relative replacement. This is achieved through controlled experiments that involve fixed and resampled assignments, layer-shared maps, and parameter-matched channel adapters. The study also includes a task-null theorem and an exact frame model to analyze the conditions under which reliance on learned transport is necessary.
Results
The experiments reveal that all tested DNSD checkpoints exhibit reliance on learned transport, but only the Roman-Empire benchmark retains a performance advantage over control models after retraining. The findings indicate that learned maps can govern computations without being indispensable, as retraining can recover performance in most cases.
Implications
The results suggest that researchers should be cautious in interpreting the significance of learned transport in GNNs. The framework developed in this study can guide future research in evaluating the necessity of learned structures in machine learning models, particularly in graph-based contexts.
Enhancing Automated Machine Learning via Homogeneous Train-Test Splitting Methods
Optimization
Theory
Efficient ML
- Standard random splitting methods often lead to unreliable performance estimates due to distributional shifts.
- Geometry-based splitting methods consistently underperform in maintaining multivariate distributional similarity.
- The proposed Optimised-Distribution method significantly improves statistical similarity between train and test sets.
- A systematic evaluation framework using multiple statistical tests was established to assess train-test splitting strategies.
Read more
Enhancing Automated Machine Learning via Homogeneous Train-Test Splitting Methods
Summary
This paper addresses the critical issue of accurate model evaluation in machine learning, particularly in the context of Automated Machine Learning (AutoML). The authors highlight that standard random splitting methods for train-test partitioning often violate the assumption of statistical similarity, leading to unreliable performance estimates, especially in datasets with class imbalance or spatial autocorrelation. The study systematically evaluates five established splitting strategies—random splitting, stratified sampling, Kennard–Stone, Duplex, and SPXY—across fifteen UCI benchmark datasets using statistical similarity measures such as chi-square, Kolmogorov–Smirnov, and Maximum Mean Discrepancy (MMD). The findings reveal that geometry-based methods tend to produce training sets that are distributionally dissimilar from test sets, resulting in instability in performance estimates. In contrast, the proposed Optimised-Distribution method explicitly maximizes statistical similarity and achieves the highest mean MMD similarity score of 89.0%. This research underscores the importance of considering distributional similarity in train-test splitting for enhancing the reliability of AutoML systems.
Methodology
The authors conducted a systematic comparative evaluation of five train-test splitting strategies using statistical similarity tests (chi-square, Kolmogorov–Smirnov, and MMD) across fifteen diverse UCI datasets. They developed the Optimised-Distribution method, which iteratively swaps samples to maximize statistical similarity between training and testing sets.
Results
The study found that geometry-based methods produced near-zero MMD scores, indicating poor distributional similarity, while the Optimised-Distribution method achieved a mean MMD similarity of 89.0%, outperforming all other strategies. This improvement in similarity led to more reliable performance estimates in the AutoML context.
Implications
The findings suggest that incorporating statistical similarity into train-test splitting can significantly enhance the reliability of model evaluations in AutoML systems. This could lead to better model selection and generalization in various applications, including healthcare and security.
Equilibrium Training of Energy-Based Models with Parallel Trajectory Tempering
Generative Models
Optimization
Efficient ML
- Introduction of Parallel Trajectory Tempering (PTT) for EBM training.
- PTT maintains equilibrium sampling, enhancing training stability and efficiency.
- Experimental results show PTT outperforms existing EBM training methods and state-of-the-art deep generative models.
- PTT provides direct estimates of thermalization times and accurate log-likelihoods.
Read more
Equilibrium Training of Energy-Based Models with Parallel Trajectory Tempering
Summary
This paper addresses the challenges of training Energy-Based Models (EBMs), which are valuable for generative modeling in scientific contexts but often suffer from poor Markov Chain Monte Carlo (MCMC) mixing. The authors propose a novel training algorithm called Parallel Trajectory Tempering (PTT), which maintains equilibrium sampling throughout the learning process by leveraging the continuity of the optimization path. This approach allows for stable and efficient training on complex, multimodal datasets, particularly in scenarios where data is scarce. PTT is combined with reservoir sampling and adaptive optimization, achieving a computational cost similar to Persistent Contrastive Divergence (PCD), making it a practical alternative to traditional training methods. The authors demonstrate that PTT not only provides direct estimates of thermalization times and equilibrium samples but also yields accurate log-likelihoods without significant additional costs. Experimental results show that PTT consistently outperforms existing EBM training techniques and surpasses state-of-the-art deep generative models on discrete tabular data, offering higher-quality samples and improved robustness against overfitting. The findings suggest that equilibrium maximum-likelihood training of EBMs can be both practical and computationally efficient, opening new avenues for scientific data analysis.
Methodology
The authors developed the Parallel Trajectory Tempering (PTT) algorithm, which involves exchanging replicas between neighboring models along the learning trajectory rather than relying on temperature-based methods. This approach is combined with reservoir sampling and adaptive optimization to ensure efficient equilibration and maintain equilibrium sampling throughout the training process.
Results
The experiments conducted on Restricted Boltzmann Machines demonstrated that PTT consistently outperformed existing EBM training approaches. On discrete tabular data, PTT achieved higher-quality samples and showed greater robustness to overfitting compared to state-of-the-art deep generative models.
Implications
The proposed PTT algorithm makes equilibrium maximum-likelihood training of EBMs practical and computationally efficient, which could significantly enhance the application of EBMs in scientific data analysis across various fields, including computational biology and neuroscience.
Tight Generalization Bound for AdaBoost
Theory
- Establishes a new upper bound for the generalization error of AdaBoost.
- Combines existing results with a new margin-based generalization bound for voting classifiers.
- Provides a tighter bound compared to previous works, improving understanding of AdaBoost's generalization capabilities.
- Confirms the effectiveness of AdaBoost in transforming weak learners into strong learners.
Read more
Tight Generalization Bound for AdaBoost
Summary
This paper presents a new upper bound on the generalization error of the AdaBoost algorithm, establishing that the error is Θ(d ln(nγ²/d)/nγ² + ln(1/δ)/n), where γ is the advantage of the weak learner, d is the VC-dimension of the hypothesis class, n is the sample size, and δ is the confidence parameter. The author combines existing knowledge about AdaBoost's output as a voting classifier with a novel margin-based generalization bound for voting classifiers. This work builds on previous findings, providing a tighter bound than earlier studies, which were suboptimal by a factor of ln(ln(nγ²/d))². The results indicate that AdaBoost can effectively generalize from weak learners, reinforcing its status as a robust boosting algorithm in machine learning.
Methodology
The author proves the upper bound on AdaBoost's generalization error by leveraging the known properties of voting classifiers and introducing a new margin-based generalization bound. The proof utilizes VC-theoretic techniques to derive the results, ensuring the bound is applicable to a wide range of weak learners.
Results
The paper concludes that the generalization error of AdaBoost is bounded by Θ(d ln(nγ²/d)/nγ² + ln(1/δ)/n), with the tightness of the bound corroborated by existing lower bounds. This result is significant as it clarifies the conditions under which AdaBoost maintains its effectiveness as a boosting algorithm.
Implications
The findings have implications for the design and analysis of boosting algorithms, particularly in understanding their generalization properties. This could lead to improved implementations of AdaBoost and inspire further research into boosting methods and their applications in various machine learning tasks.
Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework
Multimodal
Time Series
Efficient ML
- Introduces a robust multi-modal sensing platform integrating ECG and ear-PPG with IMUs to mitigate motion artifacts.
- Demonstrates that BP-related information can be effectively extracted from single beats, eliminating the need for extended temporal context.
- Presents a lightweight hybrid learning model that combines CNN-derived embeddings with physiological features, achieving a 28.2% reduction in MAE.
- Validates the framework through a multi-phase stress protocol and subject-disjoint validation, ensuring robustness in real-world conditions.
Read more
Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework
Summary
This paper addresses the challenges of continuous cuffless blood pressure (BP) monitoring, which is often hindered by motion artifacts and physiological variability. Traditional methods typically rely on multi-second windows for BP estimation, which can be problematic in real-world scenarios where signal corruption occurs. The authors propose a novel lightweight multi-modal wearable framework that integrates synchronized electrocardiography (ECG) and ear-clip photoplethysmography (PPG) sensors, each equipped with a 6-axis inertial measurement unit (IMU) to provide motion context. The framework utilizes a hybrid learning architecture combining a one-dimensional convolutional neural network (CNN) to extract beat-level features from PPG signals and a LightGBM regressor that incorporates additional physiological features such as pulse transit time (PTT) statistics and heart rate variability (HRV). The approach enables accurate BP estimation from individual beats without the need for long temporal contexts, making it suitable for wearable applications. The model was validated using a multi-phase stress protocol and the PulseDB dataset, achieving significant reductions in mean absolute error (MAE) for both systolic and diastolic BP compared to baseline models. This work demonstrates the feasibility of efficient cuffless BP monitoring under practical conditions, paving the way for improved hypertension management.
Methodology
The methodology involves a multi-modal wearable system that synchronizes ECG and ear-PPG sensors, each with an IMU. A one-dimensional CNN is used to extract features from individual PPG beats, which are then fused with physiological features (PTT and HRV) in a LightGBM regression model. This approach allows for single-beat BP estimation without relying on long temporal contexts.
Results
The proposed model achieved mean absolute errors of 4.02 ± 0.21 mmHg for systolic BP and 1.79 ± 0.05 mmHg for diastolic BP across 30 independent runs, representing a 28.2% reduction in combined MAE compared to baseline models.
Implications
The findings suggest that this lightweight framework can enhance continuous BP monitoring in wearable devices, potentially improving hypertension management and patient adherence to monitoring protocols. It opens avenues for further research into real-time health monitoring systems that are less invasive and more user-friendly.
SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
Reinforcement Learning
Large Language Models
Robotics
- SkillRise enables cross-task skill learning by organizing tasks into sequences and using a single policy for task solving and skill curation.
- The framework employs decoupled credit assignment to optimize task solving and skill curation separately.
- SkillRise achieves superior performance compared to existing methods, with significant improvements in task success rates.
- The model demonstrates cross-task scaling, improving performance as it encounters longer sequences of related tasks.
Read more
SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
Summary
The paper introduces SkillRise, a novel reinforcement learning framework designed to enhance the learning of transferable skills across related yet distinct tasks. Traditional reinforcement learning approaches treat tasks as independent episodes, which limits the ability of agents to leverage past experiences. SkillRise addresses this by organizing tasks into progressively challenging sequences, allowing a single policy to alternate between solving tasks and curating an evolving skill document. This framework employs decoupled credit assignment, where task solving is evaluated based on immediate outcomes, while skill curation is guided by future task performance. The authors demonstrate SkillRise's effectiveness through experiments on three benchmarks: ALFWorld, WebShop, and ScienceWorld. The results show that SkillRise outperforms existing methods, achieving significant gains in performance and exhibiting cross-task scaling, where performance improves with longer sequences of related tasks. This indicates that SkillRise effectively reuses transferable skills rather than relying on repeated attempts of the same task, providing a more efficient training paradigm for large language model agents.
Methodology
SkillRise constructs sequences of related tasks, allowing an agent to extract and refine skills through a single policy that alternates between solving tasks and curating a skill document. It employs decoupled credit assignment to evaluate task outcomes and skill curation based on future task performance, optimizing both behaviors simultaneously.
Results
SkillRise consistently outperformed prompting-based methods and standard reinforcement learning baselines across all tested benchmarks. The performance gains ranged from 2.3 to 8.5 percentage points over the strongest baseline, with evidence of improved performance as the model encountered longer sequences of related tasks.
Implications
The findings suggest that SkillRise can significantly enhance the efficiency and effectiveness of large language model agents in solving complex tasks by enabling them to learn and reuse transferable skills across different contexts. This could have applications in various domains requiring adaptive and intelligent task-solving capabilities.
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
NLP
Large Language Models
Graph Learning
- HYSET formulates tool retrieval as a set-level problem rather than an individual tool scoring problem.
- The method captures size-dependent interactions among tools, improving the selection process.
- HYSET consistently outperforms existing retrieval methods in practical applications.
- The approach supports generalization to unseen tools and categories with minimal supervision.
Read more
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
Summary
This paper addresses the challenge of tool retrieval for large language model (LLM) agents, which increasingly rely on external tools to perform real-world tasks. Traditional retrieval methods score tools in isolation or sequentially, failing to evaluate the joint utility of a candidate set. The authors propose a novel approach called HYSET (HYperedge-based SEt-level Tool retrieval), which formulates tool retrieval as query-conditioned hyperedge prediction on a tool co-invocation hypergraph. This method allows for scoring the entire tool set as a unit, capturing size-dependent tool compatibility through cardinality-specific interactions. HYSET is designed as a pre-selection module that does not require modifications to downstream agents. Experimental results on ToolBench demonstrate that HYSET outperforms state-of-the-art baselines in both tool retrieval performance and end-to-end task success, while also supporting zero-shot and few-shot transfer to unseen tools and domains with minimal supervision.
Methodology
The authors introduce HYSET, which utilizes a hypergraph representation to model tool co-invocation patterns. It scores tool sets as a whole based on query-conditioned hyperedge predictions, allowing for a more coherent and effective retrieval process. The methodology captures interactions among tools that vary with the size of the tool set, enhancing compatibility and utility.
Results
Experiments conducted on ToolBench show that HYSET significantly improves tool retrieval performance compared to state-of-the-art methods, achieving higher recall and completeness in the retrieved tool sets. Additionally, it demonstrates strong performance in zero-shot and few-shot scenarios, indicating its robustness and adaptability to new tasks and tools.
Implications
The findings suggest that treating tool retrieval as a set-level problem can lead to more effective LLM agent performance in real-world applications. This approach could enhance the efficiency of LLMs in various domains, particularly where multiple tools are required to complete complex tasks. The ability to generalize to unseen tools and categories also opens up new avenues for deploying LLMs in diverse environments.
From Conceptual Hydrologic Models to Conceptually Interpretable Neural Networks: A Snow-Water Mass-Conserving-Perceptron Framework for Discovering Catchment-Scale Precipitation-Storage-Runoff Representations
Interpretability
Time Series
Theory
- The Mass-Conserving Perceptron (MCP) framework integrates conceptual hydrologic models with neural networks.
- The framework achieves comparable predictive performance to conventional models while using fewer parameters.
- Significant performance improvements are observed when increasing the number of states from one to two.
- Basin-specific models can balance predictive accuracy and complexity through directed-graph representations.
Read more
From Conceptual Hydrologic Models to Conceptually Interpretable Neural Networks: A Snow-Water Mass-Conserving-Perceptron Framework for Discovering Catchment-Scale Precipitation-Storage-Runoff Representations
Summary
This paper introduces the Mass-Conserving Perceptron (MCP), a novel framework that reformulates traditional conceptual hydrologic models into physically constrained, interpretable neural networks. The authors develop a snow-water MCP network and evaluate its performance across 513 river basins in the continental United States. By recasting a coupled two-state SOIL-MCP and SNOWMCP model into a mass-conserving neural network, they demonstrate that the predictive performance of the MCP framework is comparable to that of conventional recurrent neural networks. The study investigates the impact of state-information sharing within two-state architectures and explores broader single-layer networks with varying numbers of states. Results indicate that median streamflow prediction performance improves significantly when increasing from one to two states, with diminishing returns observed beyond that. The findings suggest that basin-specific MCP networks can achieve similar predictive accuracy to LSTM models while utilizing fewer parameters. The paper emphasizes the importance of identifying the necessary states and their interactions for effective hydrologic representation, proposing future research to enhance model training with additional hydrologic observations.
Methodology
The authors developed a snow-water MCP network framework, reformulating traditional hydrologic models into mass-conserving neural networks. They evaluated the framework across 513 river basins, comparing the performance of one-state, two-state, and five-state networks, and analyzed state-information sharing within the architectures.
Results
The study found that the median KGE (Kling-Gupta Efficiency) increased from 0.82 for one-state networks to 0.89 for two-state networks and 0.90 for five-state networks, indicating diminishing returns beyond two states. Basin-specific MCP and LSTM models achieved the same median KGE of 0.90, with MCP models using fewer parameters on average.
Implications
The MCP framework offers a promising approach for hydrologic modeling, providing interpretable representations that can enhance understanding of precipitation-storage-runoff dynamics. The findings suggest potential for improved hydrologic predictions by integrating additional observations into model training.
Reinforcement Learning for Code Optimization
Reinforcement Learning
Optimization
- Introduces a three-stage framework for making execution time learnable in code optimization via RL.
- Develops DMC-Optim, a dataset with a focus on both correctness and optimization tests.
- Achieves significant improvements in optimization-aware configurations, enhancing pass rates on benchmark tasks.
- Demonstrates robustness of the proposed method against timing noise and reward sparsity.
Read more
Reinforcement Learning for Code Optimization
Summary
This paper addresses the challenges of applying reinforcement learning (RL) to code optimization, particularly in achieving both correctness and execution speed. The authors identify that while RL has been successful in generating correct code, it struggles with optimization due to issues like measurement noise and reward sparsity. To tackle these issues, they propose a three-stage approach: (1) enhancing the testing framework with large optimization tests and a calibrated sandbox, (2) refining the reward structure to incorporate both correctness and execution speed, and (3) adapting the RL model to learn effectively from the noisier reward signals. The authors introduce DMC-Optim, a dataset designed for this purpose, and demonstrate that their method significantly improves the performance of code generation models, achieving higher pass rates on optimization tasks while maintaining correctness. The results indicate that their approach can lead to substantial improvements in code efficiency, even under challenging conditions.
Methodology
The methodology involves creating a structured RL environment that separates correctness from optimization tests, using a calibrated sandbox for execution time measurement, and employing an offline simulator to refine reward structures. The authors adapt the Generalized Reinforcement Policy Optimization (GRPO) to handle the challenges of sparser and noisier rewards, ensuring stability in the learning process.
Results
The proposed method led to improvements in pass rates for optimization tasks, with Qwen 2.5 7B achieving a pass rate increase from 18.0% to 31.3% and CWM 32B from 30.7% to 50.4% at stricter evaluation percentiles. In degraded conditions, the robust optimization RL showed up to 200% improvement over standard RLVR, and CWM 32B outperformed standard RLVR in median-sample speed comparisons by up to 83%. The approach also achieved about half the human rate of complexity-class improvements.
Implications
The findings suggest that RL can be effectively applied to code optimization tasks, potentially leading to more efficient code generation tools in software development. This could have significant implications for automated programming, competitive programming, and enhancing the performance of AI-driven coding assistants.
CondPSE: A Polynomial-Filtered Structural Encoder with Conditional Modulation for Graphs
Graph Learning
- CondPSE introduces a polynomial-filtered structural encoder that enhances graph expressivity.
- The encoder achieves significant improvements in structural discrimination on synthetic benchmarks.
- Performance gains in synthetic tasks do not automatically lead to better results in real-world applications.
- The polynomial filter bank is identified as a key contributor to the performance improvements.
Read more
CondPSE: A Polynomial-Filtered Structural Encoder with Conditional Modulation for Graphs
Summary
The paper introduces CondPSE, a novel learned positional and structural encoder designed to enhance the expressiveness of graph neural networks (GNNs) beyond the limitations imposed by the 1-WL test. CondPSE employs a polynomial graph filter bank to process Gaussian node probes, followed by a conditional modulation mechanism that refines the structural-response branches based on various contextual signals. The encoder is pretrained to reconstruct node-level and graph-level targets, allowing it to be frozen and utilized in downstream tasks. The authors demonstrate that CondPSE significantly improves structural discrimination on synthetic benchmarks, achieving a remarkable increase in accuracy for CSL and EXP tasks compared to the previous GPSE model. However, the performance on real-world molecular property prediction tasks shows that the advantages gained in synthetic settings do not consistently translate to improved downstream performance, indicating a complex relationship between structural discrimination and practical application. The findings suggest that while CondPSE excels in distinguishing structural patterns, its effectiveness in real-world applications may depend on the integration with downstream architectures and the alignment of pretraining targets with task-specific labels.
Methodology
CondPSE utilizes a polynomial graph filter bank applied to Gaussian node probes, followed by FiLM-style conditional modulation that incorporates cross-filter, local, and global context. The encoder is pretrained to reconstruct positional and structural targets, allowing it to generate effective encodings for downstream graph models.
Results
CondPSE achieved CSL accuracy improvements from 42.9% to 97.3% and EXP accuracy from 68.3% to 99.9% compared to GPSE on synthetic benchmarks. However, in real molecular property prediction tasks, its performance was comparable to GPSE, indicating that strong structural discrimination does not guarantee superior downstream transfer.
Implications
The findings suggest that while advanced structural encoders like CondPSE can enhance graph representation capabilities, their practical utility in real-world applications may require careful consideration of how these encodings are integrated into downstream models. This highlights the need for further research into the alignment of pretraining objectives with specific application tasks.
Bridging Compute- and Data-Optimal Pretraining
Large Language Models
Theory
Efficient ML
- Introduction of Compute-Data (CD) scaling laws bridging compute and data optimal regimes.
- Token effectiveness function η quantifies the value of derived tokens relative to fresh tokens.
- Three operational regimes identified: compute-bound, data-bound, and model-bound.
- Diminishing returns observed when substituting compute for data as model size and data availability increase.
Read more
Bridging Compute- and Data-Optimal Pretraining
Summary
This paper introduces Compute-Data (CD) scaling laws, a novel framework that reconciles compute-optimal and data-optimal pretraining regimes in machine learning. Traditional scaling laws assume an unlimited supply of fresh data, which is increasingly unrealistic as compute resources grow faster than the availability of high-quality data. The authors propose an effectiveness function, η, that quantifies the value of derived tokens (e.g., from multi-epoch repetition or paraphrasing) relative to fresh tokens. Through extensive experiments on the Dolma-3 corpus, involving model sizes from 14M to 600M parameters, the authors demonstrate that the effectiveness of derived tokens is not constant but varies based on model size, tokens-per-parameter ratio, and the amount of derived data. The findings reveal that substituting compute for data yields diminishing returns as model size and data availability increase, leading to the identification of three operational regimes: compute-bound, data-bound, and model-bound. The CD scaling laws provide practical guidance for optimizing compute and data allocation in pretraining, suggesting when to switch between data expansion strategies and how to allocate resources effectively.
Methodology
The authors fitted the CD scaling laws by conducting experiments with various model sizes (14M to 600M parameters) and fresh data sizes (30M to 30B tokens), while exploring derived-token budgets (30M to 120B tokens). They employed hyperparameter tuning over learning rates and weight decay to analyze the effectiveness function η across different data expansion strategies, specifically multi-epoch repetition and paraphrasing.
Results
The study found that the effective-token count saturates as derived tokens increase, with both the saturation level and the rate of saturation being dependent on model size and tokens-per-parameter ratio. The fitted effectiveness function η exhibited a power-law decay, indicating that the effectiveness of derived tokens diminishes with increasing model size and data availability. The CD scaling laws successfully interpolate between compute-optimal and data-optimal scaling, providing a comprehensive understanding of pretraining dynamics.
Implications
The findings have significant implications for practitioners in machine learning, particularly in optimizing pretraining processes. The CD scaling laws can guide decisions on resource allocation between compute and data, inform the choice of data expansion strategies, and help identify the limiting factors in pretraining runs. This framework can enhance the efficiency of model training in scenarios where high-quality data is scarce.
MemSFT: Mitigating Alignment Tax with an External Parametric Memory
Large Language Models
NLP
- MemSFT effectively mitigates alignment tax by using an external parametric memory.
- The approach allows for domain specialization without compromising general model capabilities.
- MemSFT shows consistent improvements in domain performance across various LLM sizes.
- The method retains general capabilities while enhancing domain-specific knowledge.
Read more
MemSFT: Mitigating Alignment Tax with an External Parametric Memory
Summary
The paper introduces MemSFT, a novel approach to mitigate the alignment tax encountered when adapting Large Language Models (LLMs) to specialized domains. Traditional fine-tuning methods often lead to catastrophic forgetting, degrading general task performance while enhancing domain-specific capabilities. MemSFT addresses this issue by employing an external parametric memory that decouples domain specialization from the backbone model's parameters. This memory is trained to mimic a non-parametric retriever, allowing it to store and recall domain-specific knowledge without altering the backbone LLM. During inference, a learned router dynamically integrates outputs from both the memory and the backbone model, enabling selective invocation of domain expertise. The authors demonstrate the effectiveness of MemSFT across various domains, including biology, geoscience, and law, using models of different sizes. The results indicate that MemSFT significantly enhances domain performance while maintaining general capabilities, contrasting with the severe forgetting observed in traditional supervised fine-tuning (SFT) methods.
Methodology
MemSFT utilizes an external parametric memory trained to replicate the behavior of a non-parametric retriever. This memory is integrated with a frozen backbone LLM, allowing for the selective fusion of outputs during inference through a learned router. The memory is trained on domain-specific data, enabling it to internalize relevant knowledge while preserving the backbone model's general capabilities.
Results
Experimental evaluations demonstrate that MemSFT significantly improves domain-specific performance in tasks related to biology, geoscience, and law, with negligible degradation in general capabilities. In contrast, traditional SFT methods exhibit severe forgetting of general tasks, highlighting the effectiveness of the MemSFT approach.
Implications
MemSFT provides a practical solution for equipping LLMs with specialized capabilities in various domains without sacrificing their general performance. This could enhance the deployment of LLMs in professional fields requiring high reliability and expertise, such as healthcare, legal analysis, and scientific research.
Guiding Posterior Exploration with Optimizer-Derived Geometry
Optimization
Efficient ML
Theory
- Introduces a preconditioned sampling strategy using optimizer-derived geometries.
- Reduces the need for lengthy sampling burn-in phases in Bayesian neural networks.
- Maintains or improves predictive performance and uncertainty quantification.
- Validates the approach across multiple datasets and network architectures.
Read more
Guiding Posterior Exploration with Optimizer-Derived Geometry
Summary
This paper addresses the challenges of uncertainty quantification in Bayesian neural networks (BNNs) by proposing a novel method that integrates optimization and sampling phases more effectively. Traditional sampling methods, such as Markov chain Monte Carlo (MCMC), struggle with the computational cost of exploring high-dimensional and multimodal posterior distributions. The authors introduce a preconditioned sampling strategy that leverages curvature estimates derived from adaptive optimizers like AdamW during the optimization phase. This approach allows for a more informed initialization of sampling processes, significantly reducing the need for a lengthy burn-in phase and enhancing numerical stability. The proposed method maintains or improves predictive performance and uncertainty quantification without incurring additional computational costs. The authors validate their findings across various datasets and network architectures, demonstrating the robustness and efficiency of their approach in Bayesian Deep Ensembles (BDEs).
Methodology
The authors utilize curvature estimates from adaptive optimizers like AdamW to inform the preconditioning of stochastic gradient MCMC (SGMCMC) samplers. They derive several strategies for transferring optimizer properties to enhance sampling efficiency and evaluate these strategies across different modalities and architectures.
Results
The proposed method consistently shows improved efficiency in posterior exploration, with reduced burn-in times and enhanced numerical stability. It also maintains or enhances predictive performance and uncertainty quantification compared to traditional methods.
Implications
This work has significant implications for the deployment of Bayesian neural networks in high-stakes applications, where uncertainty quantification is critical. The integration of optimization and sampling phases can lead to more efficient and reliable models in various domains.
Collaborative Weighting with Pessimistic Critic for Mitigating Overestimation in Off-Policy Reinforcement Learning
Reinforcement Learning
Robotics
- Introduction of Collaborative Weighting Actor-Critic (CWAC) framework to address overestimation in off-policy RL.
- Incorporation of a collaborative weighting mechanism that adapts to TD-errors and predictive uncertainty.
- Development of a stochastic pessimistic value estimation scheme to reduce error propagation.
- CWAC shows improved performance and stability across multiple continuous control tasks.
Read more
Collaborative Weighting with Pessimistic Critic for Mitigating Overestimation in Off-Policy Reinforcement Learning
Summary
This paper addresses the challenge of overestimation bias in off-policy reinforcement learning (RL) algorithms, particularly in continuous control tasks. The authors introduce a novel framework called Collaborative Weighting Actor-Critic (CWAC), which integrates a collaborative weighting mechanism and a stochastic pessimistic value estimation scheme. The collaborative weighting mechanism accounts for predictive uncertainty in value estimation, allowing the model to prioritize reliable samples while suppressing noisy updates. The stochastic pessimistic value estimation is achieved by sampling from the return distribution, which helps mitigate error propagation during policy improvement. The proposed CWAC framework can be easily integrated into existing off-policy algorithms like Soft Actor-Critic (SAC), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Deep Deterministic Policy Gradient (DDPG) with minimal overhead. Empirical evaluations on various continuous control benchmarks demonstrate that CWAC significantly enhances performance, sample efficiency, and training stability compared to state-of-the-art methods.
Methodology
The CWAC framework employs a distributional critic to model return uncertainty and introduces a collaborative weighting mechanism that reweights TD-errors based on predictive uncertainty. It also implements stochastic sampling from the return distribution for policy improvement, which helps to mitigate overestimation bias and enhances learning stability.
Results
The experimental results indicate that CWAC outperforms existing state-of-the-art off-policy RL methods across a variety of continuous control tasks, demonstrating significant improvements in sample efficiency and training stability.
Implications
The findings suggest that CWAC can be effectively utilized in real-world applications requiring robust off-policy reinforcement learning, such as robotics, autonomous driving, and industrial scheduling, where training stability and sample efficiency are critical.
Skillful forecasting of offshore winds from satellite scatterometer constellations
Time Series
- Introduction of WindCastNet, a new paradigm for offshore wind forecasting using satellite data.
- Achieves 23% and 7% reduction in forecast errors at 1-hour and 2-hour lead times, respectively.
- Outperforms traditional NWP models at lead times up to 2.5 hours.
- Utilizes spatiotemporal encoding to handle irregular satellite data effectively.
Read more
Skillful forecasting of offshore winds from satellite scatterometer constellations
Summary
This paper introduces WindCastNet, a novel deep learning framework for forecasting offshore wind speeds and directions directly from satellite scatterometer data. Unlike traditional numerical weather prediction (NWP) models, which are not optimized for short-term forecasts, WindCastNet leverages spatiotemporally irregular satellite observations to provide accurate intraday wind forecasts. The model employs a partial convolutional long short-term memory (LSTM) network that effectively handles the challenges posed by irregular spatial coverage and asynchronous sampling of satellite data. Evaluated over the North Sea, WindCastNet demonstrates significant improvements in forecast accuracy, reducing root-mean-square errors by 23% and 7% at 1-hour and 2-hour lead times, respectively, compared to the state-of-the-art HARMONIE-AROME MEPS model. The results indicate that satellite scatterometer constellations can serve as a competitive source for short-term offshore wind forecasting, which is crucial for the integration of renewable energy into power systems.
Methodology
WindCastNet employs a partial convolutional LSTM architecture to predict offshore wind fields from satellite scatterometer observations. It encodes spatial observation masks and inter-observation intervals, allowing for continuous temporal representation and forecasts at arbitrary lead times. The model is trained on irregularly sampled satellite data from multiple constellations, overcoming challenges associated with non-uniform spatial coverage and variable revisit times.
Results
WindCastNet significantly outperforms the HARMONIE-AROME MEPS model, achieving a 23% reduction in root-mean-square error at 1-hour lead times and a 7% reduction at 2-hour lead times. The model also shows improved performance over persistence forecasts during the first three hours. However, forecast skill diminishes under strong-wind conditions and spatially non-uniform flow.
Implications
The successful application of WindCastNet for offshore wind forecasting opens new avenues for integrating satellite observations into renewable energy systems. This approach can enhance grid reliability, facilitate energy trading, and improve turbine control strategies, ultimately supporting the transition to sustainable energy sources. Additionally, the methodology may be applicable to other marine weather forecasting tasks, including tropical cyclone nowcasting.
Re-thinking Mammography Transfer Learning: The Dataset-Informed Transfer Learning (DITL) Framework for Breast Cancer Screening and Lesion Diagnosis
Computer Vision
- Introduction of the Dataset-Informed Transfer Learning (DITL) framework for mammography classification.
- Integration of dataset-specific difficulty signals with neighborhood-based supervision.
- Development of Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE) and Adaptive Neighborhood Representation Triplet (A-NR-Triplet) components.
- Achieved state-of-the-art performance on the VinDR-Mammo dataset and consistent improvements on smaller datasets.
Read more
Re-thinking Mammography Transfer Learning: The Dataset-Informed Transfer Learning (DITL) Framework for Breast Cancer Screening and Lesion Diagnosis
Summary
This paper addresses the challenges of enhancing classification performance in mammography, particularly in the context of both small curated datasets and large-scale clinical cohorts. The authors propose the Dataset-Informed Transfer Learning (DITL) framework, which integrates dataset-derived difficulty signals with neighborhood-based triplet supervision. DITL introduces two key components: Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE) and Adaptive Neighborhood Representation Triplet (A-NR-Triplet). A-DWCE assigns weights to samples based on the purity of their k-nearest neighbor labels, while A-NR-Triplet enforces intra-class compactness and inter-class separation with a learnable margin. This approach eliminates the need for hyperparameter tuning and heuristic weighting, resulting in a robust optimization strategy. The DITL framework was evaluated on the large-scale VinDR-Mammo dataset, achieving state-of-the-art performance in breast density classification, with significant improvements in accuracy, F1-score, and AUC. Additionally, DITL demonstrated consistent gains on smaller ROI datasets, bridging the gap between small-scale lesion analysis and large-scale density estimation. Overall, DITL establishes a scalable and generalizable framework for mammography classification, addressing the complexities of breast cancer screening and diagnosis.
Methodology
The DITL framework utilizes self-supervised learning (SSL) to extract dataset characteristics through nearest-neighbor analysis. It employs two adaptive objectives: A-DWCE for difficulty-based weighting and A-NR-Triplet for contrasting nearest same-class and farthest different-class features. This methodology allows for stable and dataset-informed training, enhancing classification performance.
Results
On the VinDR-Mammo dataset, DITL achieved significant improvements in classification metrics, including accuracy, F1-score, and AUC (p < 0.0001). The framework also provided statistically significant gains on smaller ROI datasets, demonstrating its effectiveness across different scales and tasks.
Implications
The DITL framework has the potential to improve computer-aided diagnosis systems in mammography, enhancing the efficiency and accuracy of breast cancer screening and diagnosis. Its scalability makes it applicable to diverse clinical settings, addressing the challenges posed by class imbalance and subtle imaging differences.
Entity Resolution in Practice: Lessons from a Self-Serve Pipeline
Theory
Efficient ML
Optimization
- No single matching algorithm is universally superior; multiple algorithms should be trained and evaluated.
- Precision and recall require separate strategies for optimization.
- False-positive links can lead to significant errors in entity merging, necessitating careful verification.
Read more
Entity Resolution in Practice: Lessons from a Self-Serve Pipeline
Summary
This paper discusses the development and evaluation of a self-serve entity resolution (ER) system, which was tested on six benchmarks with record sizes ranging from 864 to 5 million. The authors identify three critical lessons that emerged from their practical experiences, which are not commonly addressed in existing ER literature. First, they found that no single matching algorithm consistently outperforms others across different datasets; therefore, they recommend training multiple algorithm families and using an automated selection process to determine the best performer for each dataset. Second, they emphasize that precision and recall require distinct approaches: precision can be improved through hard rule-based vetoes, while recall necessitates a more diverse candidate retrieval strategy. Third, they highlight the risk of false-positive links, which can lead to incorrect entity merges, stressing the importance of re-verifying cross-group merges to prevent erroneous clustering. The authors aim to provide practical insights that can help practitioners avoid lengthy and unproductive experimentation in entity resolution tasks.
Methodology
The authors developed a structured YAML specification to encode matching logic, which serves as an audit trail and guides a lightweight matcher trained via distillation. They implemented a tournament-style evaluation of different matching algorithms (DeepMatcher, LightGBM, GAT) across various datasets, optimizing for both precision and recall through separate components for retrieval and matching.
Results
The evaluation revealed that DeepMatcher and LightGBM each won three out of six benchmarks under data-scarce conditions, while GAT did not win any. The findings underscored the necessity of tuning the system per dataset and highlighted the importance of distinct strategies for precision and recall.
Implications
The insights from this paper can guide practitioners in building more effective entity resolution systems, reducing the time spent on unproductive experiments and improving the accuracy of data integration tasks across various domains.