A naukri.com initiative
ML News
Arxiv
22h
134
Image Credit: Arxiv
Ken Utilization Layer: Hebbian Replay Within a Student's Ken for Adaptive Knowledge Tracing
- Researchers introduce KUL-KT, a biologically inspired architecture for knowledge tracing (KT) that combines Hebbian memory encoding with gradient-based consolidation in a scalable framework.
- KUL-KT adapts the principle of memory consolidation in neural systems for student modeling by incorporating a time-decaying Hebbian memory update for graceful forgetting and a Loss-aligned Internal Target (LIT) method for continual learning without backpropagation through time.
- The architecture comprises a fast Hebbian memory capturing learner interactions and a slower linear network consolidating recalled samples through gradient descent, allowing for few-shot personalization and natural forgetting without storing raw data or relying on large cohort training.
- In empirical testing, KUL-KT outperforms strong baselines on ten public KT benchmarks, leading to improved learner-perceived helpfulness and reduced difficulty in a classroom setting. Ablation studies confirm the importance of Hebbian decay and LIT for continual adaptation, positioning KUL-KT as a memory-efficient and input-flexible framework for personalized learning at scale.
Read Full Article
8 Likes
Arxiv
22h
250
Image Credit: Arxiv
Catastrophic Forgetting Mitigation via Discrepancy-Weighted Experience Replay
- Catastrophic forgetting is a challenge faced in cloud-edge object detection for traffic monitoring due to the loss of previously learned knowledge when adapting to new data distributions.
- Existing approaches like experience replay and visual prompts struggle to effectively prioritize historical data for optimal knowledge retention and adaptation.
- A new algorithm called ER-EMU is proposed to mitigate catastrophic forgetting by using adaptive experience replay and a novel Domain Distance Metric-based Experience Selection (DDM-ES) algorithm.
- Experiments on the Bellevue traffic video dataset show that ER-EMU consistently enhances the performance of cloud-edge object detection frameworks in dynamic traffic environments.
Read Full Article
15 Likes
Arxiv
22h
257
Image Credit: Arxiv
HistoART: Histopathology Artifact Detection and Reporting Tool
- Whole Slide Imaging (WSI) is crucial in modern cancer diagnostics for examining tissue specimens in detail.
- WSIs are prone to artifacts during slide preparation and scanning, potentially impacting downstream image analysis.
- A study proposes and evaluates three robust artifact detection approaches targeting common artifact types in WSIs.
- The foundation model-based approach (FMA) showed the highest performance in detecting artifacts, accompanied by a quality report scorecard.
Read Full Article
15 Likes
Arxiv
22h
272
Image Credit: Arxiv
A collaborative digital twin built on FAIR data and compute infrastructure
- The integration of machine learning with automated experimentation in self-driving laboratories is aimed at accelerating discovery and optimization tasks in science and engineering applications.
- A distributed self-driving laboratory (SDL) implementation has been developed on nanoHUB services for online simulation and FAIR data management to facilitate collaboration among geographically dispersed researchers.
- Collaborators can contribute raw experimental data to a shared central database, benefiting from analysis tools and machine learning models that dynamically update as new data is added.
- The approach enables sequential optimization through active learning, demonstrated in an example of finding the optimal recipe to combine food dyes for achieving a specific color target using readily available materials.
Read Full Article
16 Likes
Arxiv
22h
276
Image Credit: Arxiv
AdaDeDup: Adaptive Hybrid Data Pruning for Efficient Large-Scale Object Detection Training
- AdaDeDup is a hybrid data pruning framework designed to enhance the efficiency of training large-scale object detection models by integrating density-based pruning with model-informed feedback.
- The framework partitions data, applies initial density-based pruning, and uses a proxy model to adjust cluster-specific pruning thresholds adaptively based on the impact of pruning on losses within each cluster.
- Extensive experiments on Waymo, COCO, and nuScenes datasets using standard models show that AdaDeDup outperforms existing methods, reduces performance degradation, and maintains model performance while pruning 20% of data.
- AdaDeDup's effectiveness in improving data efficiency for large-scale model training is highlighted by achieving near-original model performance with significant data reduction.
Read Full Article
16 Likes
Arxiv
22h
358
Image Credit: Arxiv
SEZ-HARN: Self-Explainable Zero-shot Human Activity Recognition Network
- Human Activity Recognition (HAR) using data from Inertial Measurement Unit (IMU) sensors has practical applications in healthcare and assisted living environments.
- Zero-shot HAR (ZS-HAR) addresses data limitations in HAR models but lacks transparency in decision-making.
- A new IMU-based ZS-HAR model called SEZ-HARN provides explanations for decision-making by offering skeleton videos.
- SEZ-HARN achieves competitive Zero-shot recognition accuracy and realistic explanations on benchmark datasets.
Read Full Article
21 Likes
Arxiv
22h
0
Image Credit: Arxiv
Enhancing Reasoning Capabilities in SLMs with Reward Guided Dataset Distillation
- Enhancements in knowledge distillation techniques have led to improved capabilities in compressing Large Language Models into deployable Small Language Models.
- A new framework called AdvDistill, which is a reward-guided dataset distillation approach, has been proposed to address limitations in traditional distillation methods on reasoning tasks.
- AdvDistill utilizes rewards assigned by rule-based verifiers, based on multiple generations of responses from a teacher model, to train student models effectively.
- The study shows a significant enhancement in student model performance for mathematical and complex reasoning tasks, highlighting the advantages of incorporating reward mechanisms in dataset distillation.
Read Full Article
Like
Arxiv
22h
11
Image Credit: Arxiv
Estimating Correctness Without Oracles in LLM-Based Code Generation
- Generating code from natural language specifications using Large Language Models (LLMs) can lead to factually incorrect outputs known as hallucinations.
- A new measure of incorrectness called incoherence has been introduced to estimate error likelihood in LLM-generated code without requiring an existing correct implementation (oracle).
- Experiments show that the incoherence method can identify around two-thirds of incorrect programs with no false positives, providing a reliable alternative to oracle-based evaluations.
- The study indicates a strong correlation between LLM rankings based on correctness evaluations using an oracle and the incoherence measure, suggesting the effectiveness of the new approach.
Read Full Article
Like
Arxiv
22h
250
Image Credit: Arxiv
VoyagerVision: Investigating the Role of Multi-modal Information for Open-ended Learning Systems
- Research paper introduces VoyagerVision, a multi-modal model aiming to enhance open-ended learning systems using visual inputs.
- VoyagerVision utilizes screenshots to aid in creating structures within Minecraft, showcasing potential for interpreting spatial environments and broadening task capabilities.
- The model, an extension of Voyager, demonstrates an average creation of 2.75 unique structures within fifty iterations, marking progress in its open-ended potential.
- While successful in simpler building unit tests, VoyagerVision faces challenges in more complex structures, emphasizing room for growth.
Read Full Article
15 Likes
Arxiv
22h
82
Image Credit: Arxiv
Thinking About Thinking: SAGE-nano's Inverse Reasoning for Self-Aware Language Models
- SAGE-nano introduces text-based inverse reasoning, allowing Large Language Models (LLMs) to explain their own reasoning chains post-hoc.
- The approach in SAGE-nano uses a metacognitive structure to reflect back via attention processes and generate explanations of reasoning choices.
- Through testing on logical reasoning puzzles, math problems, and ethical dilemmas, SAGE-nano shows high reasoning accuracy (74.6% on AQUA-RAT) and explanation quality (92.1% human preference score).
- The work on inverse reasoning aims to enhance interpretability and reasoning performance in AI systems, contributing to AI safety, education, and scientific discovery.
Read Full Article
4 Likes
Arxiv
22h
186
Image Credit: Arxiv
BlackBoxToBlueprint: Extracting Interpretable Logic from Legacy Systems using Reinforcement Learning and Counterfactual Analysis
- Legacy software modernization is challenging due to a lack of documentation and understanding of original decision logic.
- A novel pipeline using Reinforcement Learning (RL) and counterfactual analysis is proposed to extract interpretable decision logic from legacy systems treated as black boxes.
- The approach involves using an RL agent to explore input space, identify decision boundaries, cluster counterfactual state transitions, and train decision trees to extract human-readable rules approximating the system's decision logic.
- The pipeline's effectiveness was demonstrated on dummy legacy systems with various complexities, showing successful focus on relevant boundary regions and accurate extraction of core logic, offering potential for generating specifications and test cases during legacy migration.
Read Full Article
11 Likes
Arxiv
22h
253
Image Credit: Arxiv
Holistic Artificial Intelligence in Medicine; improved performance and explainability
- Researchers introduce xHAIM (Explainable HAIM) as a framework leveraging Generative AI for improved performance and explainability in medical applications.
- xHAIM enhances prediction and explainability by identifying task-relevant patient data, generating patient summaries, improving predictive modeling, and providing clinical explanations.
- Evaluation on the HAIM-MIMIC-MM dataset shows xHAIM boosts average AUC from 79.9% to 90.3% for chest pathology and operative tasks, making AI more transparent and useful for clinicians.
- The xHAIM framework bridges the gap between AI advancements and clinical utility by transforming AI into an explainable decision support system.
Read Full Article
15 Likes
Arxiv
22h
182
Image Credit: Arxiv
Disentangled Feature Importance
- Standard methods for feature importance quantification underestimate contributions when predictors are correlated.
- Introduction of Disentangled Feature Importance (DFI) aims to address this limitation by transforming correlated features into independent latent variables using a transport map.
- DFI provides a principled decomposition of importance scores that sum to the total predictive variability for latent additive models and interaction-weighted functional ANOVA variances under arbitrary feature dependencies.
- Comprehensive semiparametric theory for DFI establishes root-n consistency and asymptotic normality of importance estimators in the latent space, achieving computational efficiency.
Read Full Article
11 Likes
Arxiv
22h
194
Image Credit: Arxiv
Room Scene Discovery and Grouping in Unstructured Vacation Rental Image Collections
- The research addresses the challenge of room scene discovery and grouping in unstructured vacation rental image collections.
- The proposed approach uses machine learning for room-type detection, overlap detection, and clustering of images to help travelers understand property layouts.
- A supervised pipeline is introduced, focused on efficiency, low latency, and sample-efficient learning for real-time and data-scarce environments.
- The models and pipeline developed in the research show strong performance in room scene discovery and grouping, surpassing existing methods like contrastive learning and clustering with pretrained embeddings.
Read Full Article
11 Likes
Arxiv
22h
306
Image Credit: Arxiv
Augmented Physics-Based Li-ion Battery Model via Adaptive Ensemble Sparse Learning and Conformal Prediction
- Accurate electrochemical models are crucial for safe and efficient lithium-ion battery operation in applications like electric vehicles and grid storage.
- A study introduces an Adaptive Ensemble Sparse Identification (AESI) framework to enhance reduced-order li-ion battery models by addressing unpredictable dynamics.
- The AESI framework combines an Extended Single Particle Model (ESPM) with an evolutionary ensemble sparse learning strategy and conformal prediction for uncertainty quantification.
- Evaluation highlights improved voltage prediction accuracy (up to 46% error reduction on unseen data) and reliable prediction intervals with high coverage ratios for ensemble models.
Read Full Article
18 Likes
For uninterrupted reading, download the app