A naukri.com initiative
ML News
Arxiv
20h
189
Image Credit: Arxiv
Concorde: Fast and Accurate CPU Performance Modeling with Compositional Analytical-ML Fusion
- Researchers have developed Concorde, a methodology for learning fast and accurate performance models of microarchitectures.
- Concorde uses compact performance distributions to predict program behavior based on different microarchitectural components.
- Experiments show that Concorde is over five orders of magnitude faster than a reference cycle-level simulator.
- It has an average Cycles-Per-Instruction (CPI) prediction error of about 2% across various benchmarks.
Read Full Article
11 Likes
Arxiv
20h
360
Image Credit: Arxiv
SupertonicTTS: Towards Highly Scalable and Efficient Text-to-Speech System
- Researchers have developed a novel text-to-speech (TTS) system called SupertonicTTS.
- SupertonicTTS aims to improve scalability and efficiency in speech synthesis.
- The system utilizes a low-dimensional latent space, temporal compression, and ConvNeXt blocks for a lightweight architecture.
- SupertonicTTS eliminates the need for grapheme-to-phoneme modules and external aligners through cross-attention for text-speech alignment.
Read Full Article
21 Likes
Arxiv
20h
66
Image Credit: Arxiv
CodeARC: Benchmarking Reasoning Capabilities of LLM Agents for Inductive Program Synthesis
- Inductive program synthesis, or programming by example, is a task that requires synthesizing functions from input-output examples that can generalize to unseen inputs.
- A new evaluation framework called CodeARC (Code Abstraction and Reasoning Challenge) has been proposed to benchmark the reasoning capabilities of large language model (LLM) agents in the context of inductive program synthesis.
- CodeARC utilizes an interactive setting where agents interact with a hidden target function, query it with new inputs, synthesize candidate functions, and iteratively refine their solutions using a differential testing oracle.
- Among the 18 evaluated models, o3-mini performs the best with a success rate of 52.7%, indicating the difficulty of the inductive program synthesis task.
Read Full Article
4 Likes
Arxiv
20h
40
Image Credit: Arxiv
Large Language Models are Unreliable for Cyber Threat Intelligence
- Large Language Models (LLMs) are unreliable for Cyber Threat Intelligence (CTI) tasks.
- An evaluation methodology was presented to test LLMs on CTI tasks using zero-shot learning, few-shot learning, and fine-tuning.
- Experiments with three state-of-the-art LLMs and a dataset of 350 threat intelligence reports revealed potential security risks in relying on LLMs for CTI.
- LLMs showed insufficient performance on real-size reports, inconsistency, and overconfidence despite the use of few-shot learning and fine-tuning.
Read Full Article
2 Likes
Arxiv
20h
200
Image Credit: Arxiv
The Challenge of Achieving Attributability in Multilingual Table-to-Text Generation with Question-Answer Blueprints
- Multilingual Table-to-Text NLG presents a challenge in achieving attributability.
- QA blueprints improve attributability in English examples, but not in the multilingual setting.
- Inaccuracies in machine translation of blueprints from English impact the training data.
- An in-depth analysis is conducted to understand the challenges.
Read Full Article
12 Likes
Arxiv
20h
312
Image Credit: Arxiv
RECALL-MM: A Multimodal Dataset of Consumer Product Recalls for Risk Analysis using Computational Methods and Large Language Models
- Product recalls provide valuable insights into potential risks and hazards within the engineering design process.
- A multimodal dataset, RECALL-MM, has been developed using data from the United States Consumer Product Safety Commission (CPSC) recalls database.
- The dataset facilitates data-driven risk assessment and can be used to identify product risks and guide safer design decisions.
- The dataset includes interactive clustering maps and leverages a large language model (LLM) to predict potential hazards based on product images.
Read Full Article
18 Likes
Arxiv
20h
334
Image Credit: Arxiv
Aurelia: Test-time Reasoning Distillation in Audio-Visual LLMs
- AURELIA is a novel actor-critic based audio-visual reasoning framework that improves the ability of AVLLMs to process complex multi-modal inputs without additional training.
- AVReasonBench is a challenging benchmark with 4500 audio-visual questions and detailed step-by-step reasoning, evaluating the reasoning skills of AVLLMs.
- Evaluation of 18 AVLLMs on AVReasonBench reveals limitations in their multi-modal reasoning capabilities.
- Using AURELIA, a relative improvement of up to 100% is achieved, highlighting the potential of reasoning-enhanced data generation for advancing AVLLMs.
Read Full Article
20 Likes
Arxiv
20h
193
Image Credit: Arxiv
Beyond Contrastive Learning: Synthetic Data Enables List-wise Training with Multiple Levels of Relevance
- Recent advancements in large language models (LLMs) have allowed the augmentation of information retrieval (IR) pipelines with synthetic data.
- The traditional training paradigm in contrastive learning with binary relevance labels and InfoNCE loss treats all documents that are not explicitly annotated as relevant equally, regardless of their actual degree of relevance.
- In this work, synthetic documents generated by open-source LLMs are used to create a fully synthetic ranking context of graduated relevance for training dense retrievers.
- Experiments show that this approach outperforms conventional training and achieves comparable performance to retrievers trained on real, labeled training documents.
Read Full Article
11 Likes
Arxiv
20h
44
Image Credit: Arxiv
FIESTA: Fisher Information-based Efficient Selective Test-time Adaptation
- Robust facial expression recognition in unconstrained, 'in-the-wild' environments remains challenging due to significant domain shifts between training and testing distributions.
- Existing test-time adaptation (TTA) approaches often require manual selection of parameters to update, resulting in suboptimal adaptation and high computational costs.
- A novel Fisher-driven selective adaptation framework is introduced, which dynamically identifies and updates only the most critical model parameters based on their importance as quantified by Fisher information.
- The proposed approach achieves a significant improvement in F1 score over the base model, while adapting only a small subset of parameters, making test-time adaptation more efficient and practical for real-world affective computing applications.
Read Full Article
2 Likes
Arxiv
20h
48
Image Credit: Arxiv
Joint Source-Environment Adaptation of Data-Driven Underwater Acoustic Source Ranging Based on Model Uncertainty
- Adapting pre-trained deep learning models to new and unknown environments is a difficult challenge in underwater acoustic localization.
- Implied uncertainty of pre-trained models is higher in environments with more mismatch between training and test data.
- A method is proposed to partition test samples into certain and uncertain sets and improve labeling for uncertain samples.
- The approach eliminates the need for labeled data from the target environment and results in significant improvements in model prediction accuracy.
Read Full Article
2 Likes
Arxiv
20h
152
Image Credit: Arxiv
A Lightweight Image Super-Resolution Transformer Trained on Low-Resolution Images Only
- A lightweight image super-resolution transformer trained on low-resolution images only has been introduced.
- Transformer architectures in single-image super-resolution (SISR) have a higher demand for training data compared to CNNs.
- This work utilizes a lightweight vision transformer model with LR-only training methods to address the LR-only SISR benchmark.
- The model outperforms state-of-the-art LR-only SISR methods and shows superior performance on benchmark datasets.
Read Full Article
9 Likes
Arxiv
20h
249
Image Credit: Arxiv
Localized Graph-Based Neural Dynamics Models for Terrain Manipulation
- This paper introduces a learning-based approach for terrain dynamics modeling and manipulation, leveraging the Graph-based Neural Dynamics (GBND) framework.
- The approach builds a large terrain graph but only identifies a small active subgraph for predicting the outcomes of robot-terrain interaction.
- A learning-based approach is introduced to identify a small region of interest (RoI) based on the robot's control inputs and the current scene.
- The proposed method is faster and achieves better overall prediction accuracy compared to the naive GBND.
Read Full Article
14 Likes
Arxiv
20h
368
Image Credit: Arxiv
Extracting Patient History from Clinical Text: A Comparative Study of Clinical Large Language Models
- Extracting medical history entities (MHEs) from clinical text helps structure free-text clinical notes into standardized EHRs.
- This study evaluates the performance of clinical large language models (cLLMs) in recognizing patient history entities.
- The cLLMs showed potential in reducing the time required for extracting MHEs.
- Fine-tuned GatorTron and GatorTronS demonstrated the highest performance in recognizing MHEs.
Read Full Article
22 Likes
Arxiv
20h
70
Image Credit: Arxiv
Reinforcement Learning for Active Matter
- Active matter refers to systems composed of self-propelled entities that consume energy to produce motion, exhibiting complex non-equilibrium dynamics.
- Reinforcement learning (RL) has emerged as a promising framework for addressing the complexities of active matter.
- This review explores the integration of RL for guiding and controlling active matter systems, focusing on optimal motion strategies for individual active particles and regulation of collective dynamics in active swarms.
- The application of RL in active matter systems can advance the understanding, manipulation, and control of active matter in various fields such as biological systems, robotics, and medical science.
Read Full Article
4 Likes
Arxiv
20h
171
Image Credit: Arxiv
SPIO: Ensemble and Selective Strategies via LLM-Based Multi-Agent Planning in Automated Data Science
- SPIO is a framework that leverages Large Language Models (LLMs) for orchestration of multi-agent planning in automated data science.
- SPIO consists of four key modules: data preprocessing, feature engineering, modeling, and hyperparameter tuning.
- Dedicated planning agents generate candidate strategies in each module, leading to comprehensive exploration.
- SPIO offers two variants: SPIO-S, which selects the best solution path determined by LLM; and SPIO-E, which ensembles the top k candidate plans for improved predictive performance.
Read Full Article
10 Likes
For uninterrupted reading, download the app