A naukri.com initiative
ML News
Arxiv
15h
351
Image Credit: Arxiv
Transfer learning from first-principles calculations to experiments with chemistry-informed domain transformation
- Simulation-to-Real (Sim2Real) transfer learning is gaining attention in materials science as a solution to the scarcity of experimental data.
- A transfer learning scheme from first-principles calculations to experiments based on chemistry-informed domain transformation is proposed.
- The proposed method efficiently bridges the simulation space (source domain) and the experimental data space (target domain) using prior knowledge of chemistry and the relationship between source and target quantities.
- The transfer learning model exhibits high accuracy and data efficiency, even with a small number of target data, helping to save the number of trials in real laboratories.
Read Full Article
21 Likes
Arxiv
15h
84
Image Credit: Arxiv
Efficient First-Order Optimization on the Pareto Set for Multi-Objective Learning under Preference Guidance
- Researchers propose an efficient first-order optimization method for multi-objective learning under preference guidance.
- The problem is framed as a semivectorial bilevel optimization problem, optimizing a pre-defined preference function with weakly Pareto optimal model parameters.
- To solve the problem, the multi-objective constraints are converted to a single-objective constraint using a merit function with an easy-to-evaluate gradient.
- The proposed method is shown to effectively find preference-guided optimal solutions in various synthetic and real-world problems.
Read Full Article
5 Likes
Arxiv
15h
98
Image Credit: Arxiv
Optimizing Humor Generation in Large Language Models: Temperature Configurations and Architectural Trade-offs
- This study analyzes 13 state-of-the-art large language models (LLMs) to evaluate their performance in generating technically relevant humor for software developers.
- The study tests different temperature settings and prompt variations, finding that 73% of models achieve peak performance at lower stochasticity settings.
- The analysis reveals significant performance variations across models, with certain architectures demonstrating 21.8% superiority over baseline systems.
- The study provides practical guidelines for model selection and configuration, highlighting the impact of temperature adjustments and architectural considerations on humor generation effectiveness.
Read Full Article
5 Likes
Arxiv
15h
190
Image Credit: Arxiv
Towards Understanding How Knowledge Evolves in Large Vision-Language Models
- A new research paper seeks to understand the evolution of knowledge in Large Vision-Language Models (LVLMs).
- The study delves into the analysis of internal knowledge at different levels, including single token probabilities, token probability distributions, and feature encodings.
- The research identifies two key nodes in knowledge evolution, namely critical layers and mutation layers, dividing the evolution process into rapid evolution, stabilization, and mutation.
- This study provides valuable insights into the underlying mechanisms of LVLMs and contributes to their further enhancement.
Read Full Article
11 Likes
Arxiv
15h
212
Image Credit: Arxiv
Machine Learning Prediction of Cardiovascular Risk in Type 1 Diabetes Mellitus Using Radiomics Features from Multimodal Retinal Images
- This study aimed to develop a machine learning (ML) algorithm for determining cardiovascular risk in type 1 diabetes mellitus patients using multimodal retinal images.
- Radiomic features from fundus retinography, optical coherence tomography (OCT), and OCT angiography (OCTA) images were extracted.
- ML models trained with radiomic features achieved AUC values of 0.79 for identifying moderate risk cases from high and very high-risk cases, and 0.73 for distinguishing between high and very high-risk cases.
- The addition of clinical variables improved AUC values, reaching 0.99 for identifying moderate risk cases and 0.95 for differentiating between high and very high-risk cases.
Read Full Article
12 Likes
Arxiv
15h
318
Image Credit: Arxiv
Multimodal Reference Visual Grounding
- Visual grounding focuses on detecting objects from images based on language expressions.
- A new task named Multimodal Reference Visual Grounding (MRVG) is introduced, where a model has access to a set of reference images of objects in a database.
- A novel method named MRVG-Net is introduced to solve the visual grounding problem, which achieves superior performance compared to the state-of-the-art LVLMs.
- The approach bridges the gap between few-shot detection and visual grounding, unlocking new capabilities for visual understanding.
Read Full Article
19 Likes
Arxiv
15h
51
Image Credit: Arxiv
DiaTool-DPO: Multi-Turn Direct Preference Optimization for Tool-Augmented Large Language Models
- DiaTool-DPO is a novel method that enhances Tool-Augmented Large Language Models' (TA-LLMs) dialogue capabilities through Direct Preference Optimization.
- DiaTool-DPO models TA-LLM interactions as a Markov Decision Process and categorizes user queries into 3 types based on their state transition trajectories.
- By introducing a specialized objective loss for dialogue control, DiaTool-DPO achieves substantial improvements over baseline in information gathering (94.8% vs. 44%) and tool call rejection (91% vs. 9.6%) while maintaining core functionality.
- DiaTool-DPO enables the development of TA-LLMs that can handle diverse real-world scenarios without requiring additional expert demonstrations or human labeling.
Read Full Article
3 Likes
Arxiv
15h
69
Image Credit: Arxiv
Processes Matter: How ML/GAI Approaches Could Support Open Qualitative Coding of Online Discourse Datasets
- Open coding, a key step in qualitative research, can be challenging with large discourse datasets.
- ML/GAI approaches for open coding have strengths and weaknesses, and can complement human coders.
- Line-by-line AI approaches are effective in identifying content-based codes.
- Human coders excel in interpreting conversational dynamics.
Read Full Article
4 Likes
Arxiv
15h
186
Image Credit: Arxiv
A Practical Synthesis of Detecting AI-Generated Textual, Visual, and Audio Content
- Advances in AI-generated content have raised concerns about misinformation, copyright infringement, security threats, and erosion of public trust.
- Detection techniques include observation-based strategies, linguistic and statistical analysis, model-based pipelines, watermarking and fingerprinting, and ensemble approaches.
- The paper highlights the importance of robustness, adaptation to improving generative architectures, and human-in-the-loop verification.
- Challenges include adversarial transformations, domain generalization, and ethical concerns.
Read Full Article
11 Likes
Arxiv
15h
91
Image Credit: Arxiv
Comparative Analysis of Deepfake Detection Models: New Approaches and Perspectives
- This research paper focuses on the comparative analysis of deepfake detection models.
- The study investigates the performance of the GenConViT model in relation to other architectures in the DeepfakeBenchmark.
- The evaluation of models was done using relevant metrics and new datasets, resulting in GenConViT exhibiting superior accuracy (93.82%) and generalization capacity.
- This research contributes to the advancement of deepfake detection techniques to combat the dissemination of false information.
Read Full Article
5 Likes
Arxiv
15h
106
Image Credit: Arxiv
How Post-Training Reshapes LLMs: A Mechanistic View on Knowledge, Truthfulness, Refusal, and Confidence
- Post-training is essential for the success of large language models (LLMs), transforming pre-trained base models into more useful and aligned post-trained models.
- This paper compares base and post-trained LLMs from four perspectives to understand how post-training affects LLMs internally.
- Findings reveal that post-training does not change factual knowledge storage locations, adapts knowledge representations from the base model, and develops new knowledge representations.
- Truthfulness can be effectively transferred for interventions, while refusal shows limited forward transferability between the base and post-trained models.
Read Full Article
6 Likes
Arxiv
15h
109
Image Credit: Arxiv
Scenario Discovery for Urban Planning: The Case of Green Urbanism and the Impact on Stress
- Urban environments significantly influence mental health outcomes, yet effective decision-making framework for optimizing urban policies for stress reduction remains underexplored.
- This study applies Scenario Discovery (SD) in urban planning to evaluate the effectiveness of urban vegetation interventions in stress reduction across different urban environments.
- Findings reveal that increased vegetation generally correlates with lower stress levels, but high-density urban environments, crowding, and individual psychological traits can reduce its effectiveness.
- The study showcases the Scenario Discovery framework as an approach for identifying robust policy pathways in urban planning.
Read Full Article
6 Likes
Arxiv
15h
124
Image Credit: Arxiv
Enhancing Air Quality Monitoring: A Brief Review of Federated Learning Advances
- Federated Learning (FL) offers a decentralized approach for air quality monitoring, enabling collaborative model training without sharing raw data.
- FL applications in air quality and environmental monitoring are effective in predicting pollutants and managing environmental data.
- Challenges of FL in this domain include communication overhead, infrastructure demands, generalizability issues, computational complexity, and security vulnerabilities.
- Future research should focus on optimizing communication protocols and reducing the frequency of updates to address challenges and enhance the applicability of FL in real-world environmental monitoring scenarios.
Read Full Article
7 Likes
Arxiv
15h
58
Image Credit: Arxiv
CoLa -- Learning to Interactively Collaborate with Large LMs
- A new research paper introduces CoLa, a self-guided learning paradigm for training automated language guides.
- CoLa aims to simulate human guides in problem-solving scenarios using large language models (LLMs).
- Empirical results show that CoLa consistently outperforms competitive approaches in various language tasks.
- Automated guides in CoLa demonstrate superior adaptability and reasoning strategies compared to human guides.
Read Full Article
3 Likes
Arxiv
15h
25
Image Credit: Arxiv
Route Recommendations for Traffic Management Under Learned Partial Driver Compliance
- A new paper proposes a route recommendation framework to mitigate congestion in traffic management systems.
- The framework takes into account partial driver compliance, which is often observed in reality.
- The approach involves computing an optimal flow, training a compliance model, and formulating a stochastic optimization problem.
- Simulations on a grid network show that the proposed approach reduces travel time compared to baseline strategies.
Read Full Article
1 Like
For uninterrupted reading, download the app