A naukri.com initiative
ML News
Arxiv
1d
232
Image Credit: Arxiv
From Occurrence to Consequence: A Comprehensive Data-driven Analysis of Building Fire Risk
- Building fires pose a persistent threat to life, property, and infrastructure.
- A data-driven framework analyzes U.S. fire risks by integrating over one million fire incident reports with various datasets.
- Vulnerable communities face higher fire risks due to socioeconomic disparities and outdated or vacant buildings.
- Targeted interventions, such as mandating fire safety systems and providing subsidies, can enhance fire prevention and protect vulnerable groups.
Read Full Article
14 Likes
Arxiv
1d
271
Image Credit: Arxiv
Validating Emergency Department Admission Predictions Based on Local Data Through MIMIC-IV
- This study validates hospital admission prediction models initially developed using a small local dataset from a Greek hospital.
- Using the comprehensive MIMIC-IV dataset, five algorithms were evaluated for prediction: Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Random Forest (RF), Recursive Partitioning and Regression Trees (RPART), and Support Vector Machines (SVM Radial).
- Random Forest (RF) demonstrated superior performance, achieving an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.9999, sensitivity of 0.9997, and specificity of 0.9999 when applied to the MIMIC-IV data.
- These findings highlight the robustness of Random Forest in handling complex datasets for admission prediction and provide actionable insights for improving Emergency Department (ED) management strategies.
Read Full Article
16 Likes
Arxiv
1d
0
Image Credit: Arxiv
Hierarchical Adaptive Expert for Multimodal Sentiment Analysis
- Multimodal sentiment analysis is crucial for understanding human emotions across various communication channels.
- A novel framework called Hierarchical Adaptive Expert for Multimodal Sentiment Analysis (HAEMSA) is proposed to address the challenge of effectively integrating modality-shared and modality-specific information.
- HAEMSA utilizes an evolutionary optimization approach, cross-modal knowledge transfer, and multi-task learning to capture both global and local modality representations.
- Extensive experiments demonstrate that HAEMSA outperforms previous methods, achieving improved accuracy and reducing mean absolute error on multiple benchmark datasets.
Read Full Article
Like
Arxiv
1d
100
Image Credit: Arxiv
Why Representation Engineering Works: A Theoretical and Empirical Study in Vision-Language Models
- Representation Engineering (RepE) is a powerful paradigm for enhancing AI transparency.
- In Vision-Language Models (VLMs), RepE can address challenges related to visual input overriding factual linguistic knowledge.
- Theoretical framework explaining stability of neural activity in VLMs is developed using principal eigenvector.
- This work transforms RepE into a structured theoretical framework, opening new directions for improving AI robustness, fairness, and transparency.
Read Full Article
6 Likes
Arxiv
1d
104
Image Credit: Arxiv
PowerGNN: A Topology-Aware Graph Neural Network for Electricity Grids
- The paper introduces a topology-aware Graph Neural Network (GNN) framework for predicting power system states in electricity grids with high renewable integration.
- The GNN model utilizes a graph-based representation of the power network, capturing both spatial and temporal correlations in system dynamics.
- It outperforms baseline approaches, achieving substantial improvements in predictive accuracy with average RMSEs of 0.13 to 0.17 across all predicted variables.
- The results highlight the potential of topology-aware learning for scalable and robust power system forecasting in future grids with high renewable penetration.
Read Full Article
6 Likes
Arxiv
1d
112
Image Credit: Arxiv
Zero-Shot LLMs in Human-in-the-Loop RL: Replacing Human Feedback for Reward Shaping
- Reinforcement learning often faces challenges with reward misalignment, where agents optimize for given rewards but fail to exhibit the desired behaviors.
- To address these issues, the proposed approach utilizes zero-shot, off-the-shelf large language models (LLMs) for reward shaping in continuous control tasks.
- The LLM-HFBF framework is introduced to identify and correct biases in human feedback while incorporating it into the reward shaping process.
- Empirical experiments demonstrate that the proposed approach reduces reliance on potentially biased human guidance and maintains high reinforcement learning performance.
Read Full Article
6 Likes
Arxiv
1d
116
Image Credit: Arxiv
A Spatial-temporal Deep Probabilistic Diffusion Model for Reliable Hail Nowcasting with Radar Echo Extrapolation
- A Spatial-temporal Deep Probabilistic Diffusion Model called SteamCast has been introduced for hail nowcasting with radar echo extrapolation.
- The model is trained on historical reanalysis archive from Yan'an Meteorological Bureau in China.
- SteamCast provides 30-minute nowcasts at 6-minute intervals for a single radar reflectivity variable, across 9 different vertical angles, with high resolution.
- The model delivers competitive results compared to other deep learning-based models for hail nowcasting.
Read Full Article
7 Likes
Arxiv
1d
120
Image Credit: Arxiv
Uncertainty Weighted Gradients for Model Calibration
- Model calibration is important for accurate predictions in deep neural networks.
- Existing loss functions, like focal loss, fail to achieve optimal calibration performance.
- A new loss framework is proposed, addressing misalignment and uncertainty estimation issues.
- Extensive experiments show the proposed method achieves state-of-the-art performance.
Read Full Article
7 Likes
Arxiv
1d
221
Image Credit: Arxiv
Harnessing Mixed Features for Imbalance Data Oversampling: Application to Bank Customers Scoring
- This study proposes an oversampling strategy called MGS-GRF for rare event detection in binary classification on tabular data.
- MGS-GRF is designed to handle mixed features (continuous and categorical variables) and exhibits coherence and association properties.
- The method uses a kernel density estimator and locally estimated full-rank covariances to generate continuous features, while categorical features are drawn from the original samples through a generalized random forest.
- Experimental results show that MGS-GRF outperforms other synthetic procedures in terms of predictive performances, as evaluated on both simulated and real-world datasets.
Read Full Article
13 Likes
Arxiv
1d
225
Image Credit: Arxiv
MoRE-LLM: Mixture of Rule Experts Guided by a Large Language Model
- The paper introduces a method called Mixture of Rule Experts guided by a Large Language Model (MoRE-LLM) to align machine learning models with human domain knowledge.
- MoRE-LLM combines a data-driven black-box model with knowledge extracted from a Large Language Model (LLM) to enable domain knowledge-aligned and transparent predictions.
- The Mixture of Rule Experts (MoRE) generates local rule-based surrogates during training and utilizes them for the classification task, while the LLM enhances domain knowledge alignment and provides context to the rules.
- The proposed method ensures interpretability and is evaluated on various datasets, comparing its performance with interpretable and non-interpretable baselines.
Read Full Article
13 Likes
Arxiv
1d
229
Image Credit: Arxiv
Reasoning Beyond Limits: Advances and Open Problems for LLMs
- Recent generative reasoning breakthroughs have transformed how large language models (LLMs) tackle complex problems.
- Techniques such as inference-time scaling, reinforcement learning, supervised fine-tuning, and distillation have improved reasoning capabilities of LLM models.
- A comprehensive analysis of the top 27 LLM models released between 2023 and 2025 is presented.
- Key challenges in advancing LLM capabilities, including improving multi-step reasoning, overcoming limitations in chained tasks, and enhancing long-context retrieval are discussed.
Read Full Article
13 Likes
Arxiv
1d
232
Image Credit: Arxiv
RBFleX-NAS: Training-Free Neural Architecture Search Using Radial Basis Function Kernel and Hyperparameter Detection
- RBFleX-NAS is a training-free Neural Architecture Search (NAS) framework that utilizes the Radial Basis Function (RBF) kernel and a detection algorithm.
- Conventional NAS techniques require extensive training for evaluating candidate networks.
- RBFleX-NAS outperforms other training-free NAS methods in terms of top-1 accuracy and search time.
- RBFleX-NAS introduces NAFBee, an extended activation design space, for improved activation function exploration.
Read Full Article
14 Likes
Arxiv
1d
236
Image Credit: Arxiv
A Methodology to extract Geo-Referenced Standard Routes from AIS Data
- This study proposes a methodology to analyze route between maritime points of interest and extract geo-referenced standard routes from raw AIS data.
- The methodology involves segmenting AIS data into distinct routes using a finite state machine (FSM) and aggregating the segments based on departure and destination ports.
- Iterative density-based clustering is used to model the routes, with clustering parameters assigned manually and extended to the entire dataset using linear regression.
- The unsupervised approach has been tested on a six-year AIS dataset covering the Arctic region and the Europe, Middle East, North Africa areas, proving effective in extracting standard routes with less than 5% outliers.
Read Full Article
14 Likes
Arxiv
1d
244
Image Credit: Arxiv
Cyborg Data: Merging Human with AI Generated Training Data
- Automated scoring systems used in large-scale assessment traditionally require a large quantity of hand-scored data for accurate predictions.
- Generative Large Language Models can generalize to new tasks with little to no data but still need fine-tuning.
- The proposed model distillation pipeline, named 'Cyborg Data', combines human and machine-scored responses in training.
- Student models trained on 'Cyborg Data' achieve performance similar to training on the entire dataset, using only 10% of the original hand-scored data.
Read Full Article
14 Likes
Arxiv
1d
252
Image Credit: Arxiv
ShieldAgent: Shielding Agents via Verifiable Safety Policy Reasoning
- ShieldAgent is a guardrail agent designed to enforce safety policy compliance for other autonomous agents.
- It constructs a safety policy model by extracting verifiable rules from policy documents and generates a shielding plan.
- ShieldAgent-Bench, a dataset with 3K safety-related pairs of agent instructions and action trajectories, is introduced.
- Experiments show that ShieldAgent outperforms prior methods, achieving high precision and efficiency in safeguarding agents.
Read Full Article
15 Likes
For uninterrupted reading, download the app