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

A naukri.com initiative

New

Home

ML News

Arxiv

350

Image Credit: Arxiv

SORREL: Suboptimal-Demonstration-Guided Reinforcement Learning for Learning to Branch

Mixed Integer Linear Program (MILP) solvers heavily rely on hand-crafted heuristics for branching.
Data-driven approaches have been used to automatically learn these heuristics.
Suboptimal-Demonstration-Guided Reinforcement Learning (SORREL) is proposed to learn branching using suboptimal demonstrations.
SORREL shows advanced performance in branching quality and training efficiency for various MILPs.

Read Full Article

21 Likes

Arxiv

334

Image Credit: Arxiv

FedRLHF: A Convergence-Guaranteed Federated Framework for Privacy-Preserving and Personalized RLHF

FedRLHF is a decentralized framework for Reinforcement Learning with Human Feedback (RLHF).
It addresses privacy concerns by enabling collaborative policy learning without sharing raw data or human feedback.
The framework utilizes federated reinforcement learning, allowing each client to integrate human feedback locally into their reward functions.
Empirical evaluations demonstrate that FedRLHF preserves user privacy, achieves performance similar to centralized RLHF, and enhances personalization across different client environments.

Read Full Article

20 Likes

Arxiv

107

Image Credit: Arxiv

AutoRank: MCDA Based Rank Personalization for LoRA-Enabled Distributed Learning

As data volumes expand rapidly, distributed machine learning has become essential for addressing the growing computational demands of modern AI systems. Low-Rank Adaptation (LoRA) offers a promising solution to this problem by personalizing low-rank updates rather than optimizing the entire model.
To address the limitation of manual configuration of the initial rank in LoRA-enabled distributed learning, the researchers propose AutoRank, an adaptive rank-setting algorithm inspired by the bias-variance trade-off. AutoRank leverages the MCDA method TOPSIS to dynamically assign local ranks based on the complexity of each participant's data.
Experimental results demonstrate that AutoRank significantly reduces computational overhead, enhances model performance, and accelerates convergence in highly heterogeneous federated learning environments. Through its strong adaptability, AutoRank offers a scalable and flexible solution for distributed machine learning.

Read Full Article

6 Likes

Arxiv

103

Image Credit: Arxiv

Architecture-Aware Learning Curve Extrapolation via Graph Ordinary Differential Equation

Learning curve extrapolation predicts neural network performance from early training epochs.
Existing methods neglect the impact of neural network architectures on learning curves.
A novel architecture-aware neural differential equation model is proposed to forecast learning curves continuously.
The model outperforms current state-of-the-art methods and pure time-series modeling approaches.

Read Full Article

6 Likes

Discover more

Arxiv

Image Credit: Arxiv

Spatial Clustering of Citizen Science Data Improves Downstream Species Distribution Models

Citizen science biodiversity data can improve ecological modeling and conservation efforts.
Imperfect detection of species during data collection is a challenge in ecological data.
Occupancy models address imperfect detection by modeling observation and habitat selection processes separately.
Spatial clustering algorithms for constructing sites yield better species distribution models.

Read Full Article

5 Likes

Arxiv

266

Image Credit: Arxiv

Continual Learning Using a Kernel-Based Method Over Foundation Models

Continual learning (CL) focuses on learning a sequence of tasks incrementally.
This paper addresses the challenges of class-incremental learning (CIL), including catastrophic forgetting and inter-task class separation.
The proposed method, Kernel Linear Discriminant Analysis (KLDA), utilizes features learned in a foundation model (FM) to overcome these challenges.
KLDA incorporates the Radial Basis Function (RBF) kernel and its Random Fourier Features (RFF) to improve feature representations and achieves comparable performance to joint training of all classes.

Read Full Article

16 Likes

Arxiv

139

Image Credit: Arxiv

A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation

Recent advancements in graph representation learning have led to a focus on dynamic graphs.
There is a need for generative models that can handle continuously changing temporal graph data.
In this work, a new approach called DG-Gen is proposed to model interactions in dynamic graphs using joint probabilities.
DG-Gen outperforms traditional methods in generating high-fidelity graphs and improving link prediction tasks.

Read Full Article

8 Likes

Arxiv

111

Image Credit: Arxiv

Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck

This study proposes a pre-trained Graph Neural Network (GNN) model on molecules without human annotations or prior knowledge.
The previous pre-training methods rely on functional groups, but this approach aims to capture graph-level distinctions.
The proposed method, called Subgraph-conditioned Graph Information Bottleneck (S-CGIB), generates well-distinguished graph-level representations and discovers functional groups.
Experiments show the superiority of the S-CGIB approach on molecule datasets across different domains.

Read Full Article

6 Likes

Arxiv

Image Credit: Arxiv

Machine Learning Techniques for Pattern Recognition in High-Dimensional Data Mining

This paper proposes a frequent pattern data mining algorithm based on support vector machine (SVM) in high-dimensional and sparse data environments.
The algorithm converts the frequent pattern mining task into a classification problem and utilizes SVM to improve accuracy and robustness.
Experimental results show that the proposed algorithm outperforms traditional models in terms of support, confidence, and lift.
Future research directions include incorporating deep learning and ensemble learning frameworks for improved scalability and adaptability.

Read Full Article

Arxiv

274

Image Credit: Arxiv

Beyond Human Data: Aligning Multimodal Large Language Models by Iterative Self-Evolution

Human preference alignment can greatly enhance Multimodal Large Language Models (MLLMs), but collecting high-quality preference data is costly.
A novel multimodal self-evolution framework is proposed to autonomously generate high-quality questions and answers using only unannotated images.
The framework incorporates an image-driven self-questioning mechanism, answer self-enhancement technique, and image content alignment loss function.
Experiments show that the framework performs competitively with methods using external information, providing a more efficient approach to MLLMs.

Read Full Article

16 Likes

Arxiv

246

Image Credit: Arxiv

Synthetic Tabular Data Generation for Imbalanced Classification: The Surprising Effectiveness of an Overlap Class

Handling imbalance in class distribution when building a classifier over tabular data has been a problem of long-standing interest.
One popular approach is augmenting the training dataset with synthetically generated data.
State-of-the-art deep generative models yield lower-quality minority examples than majority examples.
A technique of converting binary class labels to ternary class labels by introducing a class for the overlap region significantly improves the quality of generated data.

Read Full Article

14 Likes

Arxiv

179

Image Credit: Arxiv

Theory of Mixture-of-Experts for Mobile Edge Computing

Mobile edge computing (MEC) networks face the challenge of efficiently handling diverse machine learning tasks generated by mobile users.
The traditional approach of offloading tasks to the nearest available edge server can lead to overfitting or forgetting of previous tasks.
To address this, the mixture-of-experts (MoE) theory is introduced in MEC networks to improve continual learning (CL) performance.
An adaptive gating network is used to route tasks to specialized experts, reducing generalization error over time.

Read Full Article

10 Likes

Arxiv

258

Image Credit: Arxiv

Prompt-based Unifying Inference Attack on Graph Neural Networks

This paper introduces a novel Prompt-based unifying Inference Attack framework on Graph Neural Networks (GNNs), named ProIA.
ProIA retains the graph's topological information during pre-training, enhancing the background knowledge of the inference attack model.
It utilizes a unified prompt and introduces additional disentanglement factors in downstream attacks to adapt to task-relevant knowledge.
Extensive experiments show that ProIA enhances attack capabilities and demonstrates remarkable adaptability to various inference attacks.

Read Full Article

15 Likes

Arxiv

Image Credit: Arxiv

Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference

Large Language Models (LLMs) are becoming increasingly large, limiting their use in computationally constrained environments.
Researchers have proposed a novel approach to extract task-specific circuits from LLMs for faster inference.
The extracted subset of the LLM can perform a targeted task without additional training and with a small amount of data samples.
The resulting models are considerably smaller, reducing the number of parameters up to 82.77% and more interpretable using Mechanistic Interpretability (MI) techniques.

Read Full Article

1 Like

Arxiv

394

Image Credit: Arxiv

WebLLM: A High-Performance In-Browser LLM Inference Engine

Advancements in large language models (LLMs) have made on-device deployment practical.
WebLLM is an open-source JavaScript framework that enables high-performance LLM inference within web browsers.
It leverages WebGPU for GPU acceleration and WebAssembly for CPU computation.
WebLLM paves the way for locally powered LLM applications in web browsers.

Read Full Article

23 Likes

For uninterrupted reading, download the app