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

A naukri.com initiative

New

Home

ML News

Arxiv

263

Image Credit: Arxiv

Forward Learning with Differential Privacy

Differential privacy (DP) in deep learning is a critical concern for maintaining data confidentiality and model utility.
Forward-learning algorithms add noise during the forward pass to estimate gradients, providing potential natural differential privacy protection.
A new algorithm, DP-ULR, is introduced as a privatized forward-learning algorithm with differential privacy guarantees.
DP-ULR achieves competitive performance compared to traditional differential privacy training algorithms based on backpropagation.

Read Full Article

15 Likes

Arxiv

131

Image Credit: Arxiv

HERA: Hybrid Edge-cloud Resource Allocation for Cost-Efficient AI Agents

Large language models (LLMs) like GPT-4 predominantly operate in the cloud, incurring high operational costs.
The necessity of cloud-exclusive processing for AI agents is being reconsidered with the improved accuracy of local-based small language models (SLMs).
A lightweight scheduler called Adaptive Iteration-level Model Selector (AIMS) is proposed to partition AI agent's subtasks between SLM and LLM based on subtask features to maximize SLM usage and maintain accuracy.
Experimental results show that AIMS improves accuracy by up to 9.1% and increases SLM usage by up to 10.8% compared to existing approaches.

Read Full Article

7 Likes

Arxiv

Image Credit: Arxiv

MetaLoRA: Tensor-Enhanced Adaptive Low-Rank Fine-tuning

There has been a significant increase in the deployment of neural network models, presenting challenges in model adaptation and fine-tuning.
Low-Rank Adaptation (LoRA) has emerged as a promising parameter-efficient fine-tuning method.
This research proposes MetaLoRA, a novel parameter-efficient adaptation framework that integrates meta-learning principles.
MetaLoRA accurately captures task patterns by incorporating meta-learning mechanisms and dynamic parameter adjustment strategies.

Read Full Article

5 Likes

Arxiv

Image Credit: Arxiv

Efficient Near-Optimal Algorithm for Online Shortest Paths in Directed Acyclic Graphs with Bandit Feedback Against Adaptive Adversaries

In this paper, the authors propose an efficient algorithm for the online shortest path problem in directed acyclic graphs (DAGs) under bandit feedback against an adaptive adversary.
The algorithm achieves a near-minimax optimal regret bound of O(√|E|Tlog|X|) with high probability against any adaptive adversary.
The algorithm utilizes a novel loss estimator and a centroid-based decomposition to attain this regret bound.
The algorithm's application extends to various domains, including extensive-form games, shortest walks in directed graphs, hypercubes, and multi-task multi-armed bandits, providing improved regret guarantees in each of these settings.

Read Full Article

5 Likes

Discover more

Arxiv

116

Image Credit: Arxiv

Informed Greedy Algorithm for Scalable Bayesian Network Fusion via Minimum Cut Analysis

This paper presents the Greedy Min-Cut Bayesian Consensus (GMCBC) algorithm for the structural fusion of Bayesian Networks (BNs).
GMCBC integrates principles from flow network theory into BN fusion, adapting the Backward Equivalence Search (BES) phase of the Greedy Equivalence Search (GES) algorithm and applying the Ford-Fulkerson algorithm for minimum cut analysis.
Experimental results on synthetic Bayesian Networks demonstrate that GMCBC achieves near-optimal network structures.
In federated learning simulations, GMCBC produces a consensus network that improves structural accuracy and dependency preservation compared to the average of the input networks, resulting in a structure that better captures the real underlying (in)dependence relationships.

Read Full Article

7 Likes

Arxiv

207

Image Credit: Arxiv

Less is More: Efficient Black-box Attribution via Minimal Interpretable Subset Selection

Researchers propose LiMA (Less input is More faithful for Attribution), a novel black-box attribution mechanism for AI systems.
LiMA reformulates attribution of important regions as an optimization problem for submodular subset selection.
The method accurately assesses input-prediction interactions and improves optimization efficiency using a bidirectional greedy search algorithm.
Experiments show that LiMA provides faithful interpretations with fewer regions, exhibits strong generalization, and outperforms other attribution algorithms.

Read Full Article

12 Likes

Arxiv

342

Image Credit: Arxiv

Enhancing stroke disease classification through machine learning models via a novel voting system by feature selection techniques

Heart disease remains a leading cause of mortality, necessitating accurate predictive models.
Nine machine learning algorithms were applied, including XGBoost, logistic regression, decision tree, random forest, KNN, SVM, NB Gaussian, adaptive boosting, and linear regression.
Feature selection techniques were used to refine the models and enhance performance and interpretability.
XGBoost demonstrated exceptional performance with 99% accuracy, precision, F1-score, 98% recall, and 100% ROC AUC.

Read Full Article

20 Likes

Arxiv

Image Credit: Arxiv

ParallelFlow: Parallelizing Linear Transformers via Flow Discretization

Researchers introduce a theoretical framework called Parallel Flows for analyzing linear attention models using matrix-valued state space models (SSMs).
The approach of Parallel Flows decouples temporal dynamics from implementation constraints, allowing independent analysis of chunking, parallelization, and information aggregation.
The framework reinterprets chunking procedures as computations of the flows governing system dynamics, connecting it to mathematical tools from rough path theory.
The application of Parallel Flows to DeltaNet in a low-rank setting allows for the design of simple, streamlined generalizations with lower complexity, demonstrating the power of theoretical analysis in inspiring new computational approaches.

Read Full Article

4 Likes

Arxiv

112

Image Credit: Arxiv

Operator Learning with Domain Decomposition for Geometry Generalization in PDE Solving

Neural operators have gained popularity in solving partial differential equations (PDEs) due to their ability to capture complex mappings in function spaces over complex domains.
The data requirements of neural operators limit their widespread use and transferability to new geometries.
To overcome this issue, a local-to-global framework called operator learning with domain decomposition is proposed for solving PDEs on arbitrary geometries.
The framework utilizes an iterative scheme called Schwarz Neural Inference (SNI) to solve local problems with neural operators and stitch local solutions to construct a global solution.

Read Full Article

6 Likes

Arxiv

131

Image Credit: Arxiv

Training Frozen Feature Pyramid DINOv2 for Eyelid Measurements with Infinite Encoding and Orthogonal Regularization

Accurate measurement of eyelid parameters such as MRD1, MRD2, and LF is limited by manual methods.
Deep learning models, including DINOv2, are evaluated for automating these measurements using smartphone-acquired images.
DINOv2, pretrained through self-supervised learning, demonstrates scalability and robustness, especially under frozen conditions ideal for mobile deployment.
Enhancements such as focal loss, orthogonal regularization, and binary encoding strategies improve generalization and prediction accuracy of DINOv2.

Read Full Article

7 Likes

Arxiv

Image Credit: Arxiv

Adversarial Curriculum Graph-Free Knowledge Distillation for Graph Neural Networks

Researchers propose a new method called Adversarial Curriculum Graph-Free Knowledge Distillation (ACGKD) for data-free knowledge distillation of graph neural networks.
ACGKD leverages the Binary Concrete distribution to model graph structures and introduces a spatial complexity tuning parameter, reducing the spatial complexity of pseudo-graphs.
The proposed method accelerates the distillation process by enabling efficient gradient computation for the graph structure.
ACGKD achieves state-of-the-art performance in distilling knowledge from GNNs without training data.

Read Full Article

5 Likes

Arxiv

335

Image Credit: Arxiv

Geometric Median Matching for Robust k-Subset Selection from Noisy Data

Data pruning is crucial for selecting a representative subset from a large dataset for deep learning models.
Geometric Median (GM) Matching is a novel k-subset selection strategy that leverages robust estimation to enhance resilience against noisy data.
GM Matching enjoys improved convergence rate and outperforms existing pruning approaches in high-corruption settings and high pruning rates.
It is considered a strong baseline for robust data pruning according to extensive experiments.

Read Full Article

20 Likes

Arxiv

301

Image Credit: Arxiv

NeuraLUT-Assemble: Hardware-aware Assembling of Sub-Neural Networks for Efficient LUT Inference

Efficient neural networks (NNs) using lookup tables (LUTs) have shown potential for AI applications on FPGAs.
NeuraLUT-Assemble is a framework that combines mixed-precision techniques and assembly of larger neurons to improve accuracy and connectivity of LUT-based designs.
It introduces skip-connections to enhance gradient flow and achieves competitive accuracy in various tasks.
NeuraLUT-Assemble also demonstrates up to 8.42x reduction in the area-delay product compared to the state-of-the-art at the time of publication.

Read Full Article

18 Likes

Arxiv

237

Image Credit: Arxiv

Data Cleansing for GANs

A new approach is proposed for improving the performance of generative adversarial networks (GANs) by identifying and removing harmful training instances.
The challenge with previous approaches is that they are not easily applicable to GANs due to the indirect effect of training instances on GAN parameters.
The proposed approach uses the Jacobian of the generator's gradient with respect to the discriminator's parameters to estimate the influence of instances.
By removing the identified harmful instances, the generative performance of GANs is significantly improved on various evaluation metrics.

Read Full Article

14 Likes

Arxiv

Image Credit: Arxiv

Feature Subset Weighting for Distance-based Supervised Learning through Choquet Integration

This paper introduces feature subset weighting using monotone measures for distance-based supervised learning.
The proposed method incorporates feature weights using the Choquet integral, enabling the distances to capture non-linear relationships and interactions among attributes.
An advantage of this approach is that the computed subset weights are computationally feasible, reducing the number of calculations compared to calculating all feature subset weights.
Experimental evaluation demonstrates the effectiveness of the proposed distance measure in a k-nearest neighbors classification setting.

Read Full Article

5 Likes

For uninterrupted reading, download the app