A naukri.com initiative
ML News
Arxiv
1w
186
Image Credit: Arxiv
Generative AI Enhanced Financial Risk Management Information Retrieval
- This paper introduces RiskData, a dataset curated for enhancing embedding models in risk management.
- They also introduce RiskEmbed, a finetuned embedding model designed to improve retrieval accuracy in financial question-answering systems.
- The dataset is derived from 94 regulatory guidelines published by the Office of the Superintendent of Financial Institutions from 1991 to 2024.
- Experimental results show that RiskEmbed outperforms general-purpose and financial embedding models, achieving significant improvements in ranking metrics.
Read Full Article
11 Likes
Arxiv
1w
211
Image Credit: Arxiv
Going beyond explainability in multi-modal stroke outcome prediction models
- This study aims to enhance interpretability and explainability of multi-modal prediction models integrating imaging and tabular patient data.
- The adapted xAI methods were used to generate explanation maps for identification of relevant image features and error analysis.
- The dTMs achieve state-of-the-art prediction performance, with area under the curve (AUC) values close to 0.8.
- Explanation maps calculated from brain imaging dTMs for functional outcome highlighted critical brain regions such as the frontal lobe.
Read Full Article
12 Likes
Arxiv
1w
109
Image Credit: Arxiv
On the Effectiveness and Generalization of Race Representations for Debiasing High-Stakes Decisions
- Understanding and mitigating biases in large language models (LLMs) is crucial for their use in high-stakes decision-making.
- The study introduces two decision tasks, Admissions and Hiring, to assess racial bias in LLMs.
- The experiment shows that Gemma 2B Instruct and LLaMA 3.2 3B Instruct have strong biases in favor of certain racial groups.
- While prompt engineering fails to promote fairness, debiasing interventions based on identifying 'race subspaces' within the model activations show promise in reducing biases.
Read Full Article
6 Likes
Arxiv
1w
114
Image Credit: Arxiv
Leveraging State Space Models in Long Range Genomics
- Long-range dependencies are critical for understanding genomic structure and function.
- State Space Models (SSMs) are explored as a promising alternative to conventional methods.
- SSMs match transformer performance and exhibit impressive zero-shot extrapolation across multiple tasks.
- SSMs are efficient and scalable for long-context genomic analysis and can process sequences of 1M tokens on a single GPU.
Read Full Article
6 Likes
Arxiv
1w
253
Image Credit: Arxiv
DeepGDel: Deep Learning-based Gene Deletion Prediction Framework for Growth-Coupled Production in Genome-Scale Metabolic Models
- Researchers have developed a deep learning-based gene deletion prediction framework for growth-coupled production in genome-scale metabolic models.
- The framework leverages deep learning algorithms to learn and integrate sequential gene and metabolite data representation.
- It demonstrates substantial improvements over the baseline method, with an increase in overall accuracy across different metabolic models
- The source code and examples for the framework are publicly available on GitHub.
Read Full Article
15 Likes
Arxiv
1w
359
Image Credit: Arxiv
Hybrid Temporal Differential Consistency Autoencoder for Efficient and Sustainable Anomaly Detection in Cyber-Physical Systems
- This study focuses on anomaly detection in cyber-physical systems (CPS) using an efficient and sustainable approach.
- The proposed hybrid TDC-AE approach captures the dynamics of the system by leveraging time correlations in sensor data.
- It achieves state-of-the-art classification performance, outperforming the BATADAL challenge leader without domain-specific knowledge.
- The hybrid structure maintains computational efficiency, making it broadly applicable and enhances the resilience of CPS.
Read Full Article
21 Likes
Arxiv
1w
393
Image Credit: Arxiv
A Geometric-Aware Perspective and Beyond: Hybrid Quantum-Classical Machine Learning Methods
- Geometric Machine Learning (GML) has shown improved performance by considering non-Euclidean geometry in data spaces.
- Quantum Machine Learning (QML) leverages quantum state manifolds for learning tasks.
- QML is a specialized branch of GML, with quantum states residing on curved manifolds.
- Hybrid classical-quantum architectures demonstrate tangible benefits by combining classical feature extraction with quantum embeddings.
Read Full Article
23 Likes
Arxiv
1w
359
Image Credit: Arxiv
From Broadcast to Minimap: Achieving State-of-the-Art SoccerNet Game State Reconstruction
- Game State Reconstruction (GSR) is a crucial task in Sports Video Understanding, enabling precise tracking and localization of individuals on the football field.
- A single-camera setup for GSR is challenging due to camera movements, occlusions, and dynamic scene content.
- Researchers have developed a robust end-to-end pipeline for GSR using a single-camera setup.
- The method achieved state-of-the-art game state reconstruction and outperformed other methods in the SoccerNet Game State Reconstruction Challenge 2024.
Read Full Article
21 Likes
Arxiv
1w
56
Image Credit: Arxiv
Deep spatio-temporal point processes: Advances and new directions
- Spatio-temporal point processes (STPPs) are used to model discrete events distributed in time and space.
- Traditional models often rely on parametric kernels, limiting their ability to capture heterogeneous, nonstationary dynamics.
- Recent innovations integrate deep neural architectures to model the conditional intensity function or learn flexible, data-driven influence kernels.
- The article discusses the development of the deep influence kernel approach, its components, applications in crime analysis, earthquake aftershock prediction, and sepsis prediction modeling, and promising directions for the field.
Read Full Article
3 Likes
Arxiv
1w
21
Image Credit: Arxiv
The Zero Body Problem: Probing LLM Use of Sensory Language
- Language models are being explored to approximate human use of sensory language, but they differ significantly from human usage.
- In a study using a corpus of parallel human and model responses to short story prompts, it was found that models generate stories that differ from human usage of sensory language.
- Gemini models use significantly more sensory language than humans, while models from the other families use significantly less.
- Linear probes suggest that models are capable of identifying sensory language, but instruction tuning may discourage its usage.
Read Full Article
1 Like
Arxiv
1w
46
Image Credit: Arxiv
Low Rank Learning for Offline Query Optimization
- Researchers have introduced LimeQO, a framework for offline query optimization, aiming to reduce the resource usage in learned query optimizers.
- LimeQO leverages low-rank learning to efficiently explore alternative query plans and predicts unobserved query plan latencies using purely linear methods.
- The framework models the workload as a partially observed, low-rank matrix, significantly reducing computational overhead compared to neural networks.
- LimeQO provides a low-overhead solution and a no-regressions guarantee without making assumptions about the underlying DBMS.
Read Full Article
2 Likes
Arxiv
1w
219
Image Credit: Arxiv
S'MoRE: Structural Mixture of Residual Experts for LLM Fine-tuning
- S'MoRE is a framework that integrates the efficiency of low-rank adaptations with the flexibility of MoE architectures.
- S'MoRE employs hierarchical low-rank decomposition of expert weights, resulting in residuals interconnected in a multi-layer structure.
- Residuals allow S'MoRE to emulate the capacity of many experts while instantiating and assembling just a few low-rank matrices.
- Theoretical analysis and empirical results demonstrate that S'MoRE achieves superior fine-tuning performance, offering an efficient approach for adapting LLM.
Read Full Article
13 Likes
Arxiv
1w
363
Image Credit: Arxiv
Graph Neural Network-Based Distributed Optimal Control for Linear Networked Systems: An Online Distributed Training Approach
- This paper presents a method for distributed optimal control for linear networked systems using graph recurrent neural networks (GRNNs).
- The existing approaches for this problem result in centralized optimal controllers with offline training processes.
- The main contribution of this work is the development of a distributed, online training approach for the optimal controllers.
- The method is demonstrated through numerical simulations using a dedicated simulator.
Read Full Article
21 Likes
Arxiv
1w
42
Image Credit: Arxiv
Classifying Subjective Time Perception in a Multi-robot Control Scenario Using Eye-tracking Information
- Automation and mobile robotics in work environments can lead to potential stress and cognitive overload.
- Subjective time perception is used as an indicator of well-being and cognitive strain.
- The study focuses on using eye-tracking data to estimate a person's subjective time perception in a human-swarm interaction scenario.
- The approach successfully estimates a person's time perception from eye-tracking data.
Read Full Article
2 Likes
Arxiv
1w
338
Image Credit: Arxiv
Missing Premise exacerbates Overthinking: Are Reasoning Models losing Critical Thinking Skill?
- The response length of reasoning language models (LLMs) increases for ill-posed questions with missing premises (MiP), leading to redundant and ineffective thinking.
- This scenario exacerbates the overthinking issue, which is named as MiP-Overthinking.
- LLMs not specifically trained for reasoning perform better on the MiP scenario, producing shorter responses and quickly identifying ill-posed queries.
- The current training recipe for reasoning LLMs lacks efficient thinking and encourages overthinking, indicating a critical flaw.
Read Full Article
20 Likes
For uninterrupted reading, download the app