Machine Learning (ML) Latest News and Trending articles from all top sources only on Techminis

A naukri.com initiative

New

Home

ML News

Arxiv

377

Image Credit: Arxiv

Outcome-Refining Process Supervision for Code Generation

Large Language Models have shown remarkable capabilities in code generation.
Process supervision through learned reward models is a promising approach for guiding reasoning steps, but it requires expensive training data and suffers from unreliable evaluation.
The proposed Outcome-Refining Process Supervision paradigm treats outcome refinement itself as the process to be supervised.
Experiments show that this approach enables smaller models to achieve high success accuracy and performance metrics on competitive programming tasks, without requiring training reward models.

Read Full Article

22 Likes

Arxiv

251

Image Credit: Arxiv

Adaptive Pruning for Large Language Models with Structural Importance Awareness

Researchers propose a novel pruning method called structurally-aware adaptive pruning (SAAP) for large language models (LLMs).
SAAP aims to reduce computational and memory costs while maintaining model performance on resource-constrained edge devices.
The method defines an adaptive importance fusion metric to evaluate the importance of all coupled structures in LLMs.
Experimental results show that SAAP outperforms several state-of-the-art baseline methods in terms of accuracy and token generation speed.

Read Full Article

15 Likes

Arxiv

243

Image Credit: Arxiv

Jet: A Modern Transformer-Based Normalizing Flow

Normalizing flows have emerged as a promising class of generative models for natural images.
This paper revisits the design of coupling-based normalizing flow models and uses computational blocks based on the Vision Transformer architecture.
The authors achieve state-of-the-art quantitative and qualitative performance with a simpler architecture.
While the visual quality is still behind the current state-of-the-art, strong normalizing flow models can serve as building blocks for more powerful generative models.

Read Full Article

14 Likes

Arxiv

310

Image Credit: Arxiv

Leveraging Color Channel Independence for Improved Unsupervised Object Detection

Object-centric architectures can learn to extract distinct object representations from visual scenes.
RGB color space is commonly assumed to be optimal for unsupervised learning in computer vision.
This work challenges the assumption and explores the use of other color spaces, such as HSV.
The proposed approach, using the RGB-S color space, improves reconstruction and disentanglement in object-centric representation learning.

Read Full Article

18 Likes

Discover more

Arxiv

Image Credit: Arxiv

HPC-Coder-V2: Studying Code LLMs Across Low-Resource Parallel Languages

Large Language Model (LLM) based coding tools have gained success as software development assistants, but struggle in specialized domains like high performance computing (HPC).
This study focuses on fine-tuning a specialized HPC LLM to understand the challenges of generating parallel code.
The study identifies hurdles that hold back LLMs in HPC and proposes solutions to overcome them.
Based on the findings, a specialized HPC LLM is created and evaluated, showing superior performance in open-source parallel code generation.

Read Full Article

1 Like

Arxiv

331

Image Credit: Arxiv

STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning

Robot learning is experiencing growth in the size and complexity of pre-collected datasets.
This work proposes a paradigm of training policies during deployment, retrieving and training models on relevant data at test time.
The proposed technique, STRAP, leverages pre-trained vision foundation models and dynamic time warping to retrieve sub-sequences of trajectories from large training corpora.
STRAP outperforms existing methods in terms of data utilization, generalization, and robustness in adapting policies to novel problems.

Read Full Article

19 Likes

Arxiv

305

Image Credit: Arxiv

LlamaFusion: Adapting Pretrained Language Models for Multimodal Generation

LlamaFusion is a framework that enhances pretrained text-only large language models (LLMs) with multimodal generative capabilities.
It enables LLMs to understand and generate both text and images in arbitrary sequences.
LlamaFusion utilizes dedicated modules for processing text and images, allowing interactions between text and image features.
Through experiments, LlamaFusion shows improved image understanding and generation while maintaining the language capabilities of text-only LLMs.

Read Full Article

18 Likes

Arxiv

188

Image Credit: Arxiv

Preventing Local Pitfalls in Vector Quantization via Optimal Transport

Vector-quantized networks (VQNs) have shown great performance but suffer from training instability.
Researchers propose OptVQ, a vector quantization method that integrates optimal transport to improve stability and efficiency of training.
OptVQ uses the Sinkhorn algorithm to optimize the optimal transport problem.
Experiments demonstrate that OptVQ achieves 100% codebook utilization and outperforms current state-of-the-art VQNs in image reconstruction quality.

Read Full Article

11 Likes

Arxiv

171

Image Credit: Arxiv

LiDAR-RT: Gaussian-based Ray Tracing for Dynamic LiDAR Re-simulation

This paper introduces LiDAR-RT, a framework for real-time LiDAR re-simulation in dynamic driving scenarios.
LiDAR-RT overcomes the limitations of previous methods by using Gaussian primitives and hardware-accelerated ray tracing technology.
The framework models the physical properties of LiDAR sensors using Gaussian primitives and incorporates scene graphs to handle scene dynamics.
LiDAR-RT outperforms existing methods in terms of rendering quality and efficiency.

Read Full Article

10 Likes

Arxiv

259

Image Credit: Arxiv

AutoTrust: Benchmarking Trustworthiness in Large Vision Language Models for Autonomous Driving

Recent advancements in large vision language models (VLMs) tailored for autonomous driving (AD) have shown strong scene understanding and reasoning capabilities.
AutoTrust is introduced as a comprehensive trustworthiness benchmark for large vision-language models in autonomous driving (DriveVLMs), considering trustfulness, safety, robustness, privacy, and fairness.
The benchmark includes evaluations of six publicly available VLMs and identifies vulnerabilities in DriveVLMs related to trustworthiness threats, sensitive information disclosure, adversarial attacks, and unbiased decision-making.
Immediate and decisive action is needed to address the trustworthiness of DriveVLMs, ensuring public safety in autonomous transportation systems.

Read Full Article

15 Likes

Arxiv

251

Image Credit: Arxiv

OpenEMMA: Open-Source Multimodal Model for End-to-End Autonomous Driving

OpenEMMA is an open-source end-to-end framework based on Multimodal Large Language Models (MLLMs) for Autonomous Driving (AD).
OpenEMMA incorporates the Chain-of-Thought reasoning process and achieves significant improvements compared to the baseline.
It offers effectiveness, generalizability, and robustness across a variety of challenging driving scenarios.
The code for OpenEMMA is released on GitHub: https://github.com/taco-group/OpenEMMA.

Read Full Article

15 Likes

Arxiv

268

Image Credit: Arxiv

Selective Uncertainty Propagation in Offline RL

Researchers have proposed a method called selective uncertainty propagation for confidence interval construction in offline reinforcement learning.
The method is designed to address the challenges of estimating treatment effects and dealing with distributional shifts in real-world RL instances.
Selective uncertainty propagation adapts to the level of difficulty associated with distribution shift challenges.
The technique has shown promising results in toy environments and is beneficial for offline policy learning.

Read Full Article

16 Likes

Arxiv

Image Credit: Arxiv

Learning Discretized Neural Networks under Ricci Flow

This paper introduces a study on Discretized Neural Networks (DNNs) composed of low-precision weights and activations.
The use of Straight-Through Estimator (STE) to approximate gradients for training-based DNNs introduces gradient mismatch.
The paper proposes addressing the gradient mismatch as a metric perturbation in a Riemannian manifold through the lens of duality theory.
Experimental results demonstrate the superior and stable performance of the proposed method for DNNs compared to other training-based methods.

Read Full Article

Arxiv

Image Credit: Arxiv

Scaling Laws for Imitation Learning in Single-Agent Games

Imitation Learning (IL) is widely used in machine learning, but often fails to fully recover expert behavior in single-agent games.
This study investigates the impact of scaling up model and data size on IL performance.
The findings show that IL loss and mean return scale smoothly with compute budget, resulting in power laws for training compute-optimal agents.
NetHack agents trained with IL outperform previous state-of-the-art by 1.5x.

Read Full Article

Arxiv

155

Image Credit: Arxiv

DTW+S: Shape-based Comparison of Time-series with Ordered Local Trend

Measuring distance or similarity between time-series data is crucial in various applications.
Existing measures often fail to capture similarities among local trends and can be misleading.
Researchers propose a novel measure called DTW+S that creates an interpretable matrix representation of time-series data.
DTW+S shows promising results in clustering epidemic curves, ensemble building, and classification.

Read Full Article

9 Likes

For uninterrupted reading, download the app