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

An Information-Geometric Approach to Artificial Curiosity

Learning in environments with sparse rewards remains a fundamental challenge in reinforcement learning.
Artificial curiosity addresses this limitation through intrinsic rewards to guide exploration.
Leveraging information geometry, intrinsic rewards can be constrained to concave functions of the reciprocal occupancy.
This framework integrates foundational exploration methods into a single, cohesive model.

Read Full Article

21 Likes

Arxiv

328

Image Credit: Arxiv

SPoRt -- Safe Policy Ratio: Certified Training and Deployment of Task Policies in Model-Free RL

To apply reinforcement learning to safety-critical applications, safety guarantees during policy training and deployment are necessary.
The paper presents the concept of Safe Policy Ratio (SPoRt) to provide a bound on the probability of violating a safety property in a model-free, episodic setup.
SPoRt includes Projected PPO, a new approach for training task-specific policies while maintaining a user-specified bound on property violation.
The experimental results demonstrate the trade-off between safety guarantees and task-specific performance in SPoRt.

Read Full Article

19 Likes

Arxiv

Image Credit: Arxiv

Sharpness-Aware Parameter Selection for Machine Unlearning

The article discusses the task of removing sensitive personal information from trained machine learning models.
Proposed method focuses on identifying the subset of model parameters that have the largest contribution in the unlearning process.
The selection of these parameters and updating them during unlearning leads to improved efficacy with low computational cost.
The strategy is supported by theoretical justifications and empirical evidence.

Read Full Article

2 Likes

Arxiv

384

Image Credit: Arxiv

Understanding Machine Unlearning Through the Lens of Mode Connectivity

Machine Unlearning aims to remove undesired information from trained models without requiring full retraining from scratch.
This paper investigates and analyzes machine unlearning through the lens of mode connectivity, which refers to the phenomenon where independently trained models can be connected by smooth low-loss paths in the parameter space.
The study explores mode connectivity in unlearning across various conditions, including different unlearning methods, models trained with and without curriculum learning, and models optimized with different techniques.
The findings reveal patterns of fluctuation in evaluation metrics and highlight the (dis)similarity between unlearning methods.

Read Full Article

23 Likes

Discover more

Arxiv

Image Credit: Arxiv

PEEL the Layers and Find Yourself: Revisiting Inference-time Data Leakage for Residual Neural Networks

This paper investigates the risks of inference-time data leakage in deep neural networks (NNs).
The study focuses on residual NNs, specifically the use of skip connections in residual blocks.
The authors propose a backward feature inversion method called PEEL to recover input features from intermediate outputs of residual NNs.
PEEL surpasses state-of-the-art recovery methods in terms of mean squared error (MSE) when tested on facial image datasets and pre-trained classifiers.

Read Full Article

3 Likes

Arxiv

Image Credit: Arxiv

Releasing Differentially Private Event Logs Using Generative Models

Researchers have introduced two novel approaches for releasing differentially private trace variants based on trained generative models.
The first approach, called TraVaG, utilizes Generative Adversarial Networks (GANs) to sample from a privatized implicit variant distribution.
The second approach employs Denoising Diffusion Probabilistic Models that reconstruct artificial trace variants from noise via trained Markov chains.
Both methods offer industry-scale benefits and elevate the degree of privacy assurances, particularly in scenarios featuring a substantial prevalence of infrequent variants.

Read Full Article

5 Likes

Arxiv

201

Image Credit: Arxiv

Federated Neural Architecture Search with Model-Agnostic Meta Learning

Federated Neural Architecture Search (NAS) enables collaborative search for optimal model architectures tailored to heterogeneous data to achieve higher accuracy.
FedMetaNAS is a framework that integrates meta-learning with NAS in the context of Federated Learning (FL) to accelerate architecture search by pruning the search space and eliminating the retraining stage.
It utilizes the Gumbel-Softmax reparameterization and Model-Agnostic Meta-Learning techniques to facilitate relaxation of mixed operations and adapt weights and architecture parameters for individual tasks.
Experimental evaluations demonstrate that FedMetaNAS significantly accelerates the search process by more than 50% with higher accuracy compared to FedNAS.

Read Full Article

11 Likes

Arxiv

274

Image Credit: Arxiv

Exploiting Meta-Learning-based Poisoning Attacks for Graph Link Prediction

Link prediction in graph data utilizes various algorithms and machine learning/deep learning models to predict potential relationships between graph nodes.
Recent research highlights the vulnerability of link prediction models to adversarial attacks, such as poisoning and evasion attacks.
A new approach using meta-learning techniques is proposed to exploit Variational Graph Auto-Encoder (VGAE) model's link prediction performance.
Comprehensive experiments demonstrate that the proposed approach significantly diminishes link prediction performance and outperforms other state-of-the-art methods.

Read Full Article

16 Likes

Arxiv

325

Image Credit: Arxiv

GTS-LUM: Reshaping User Behavior Modeling with LLMs in Telecommunications Industry

GTS-LUM is a user behavior model that reshapes modeling paradigms in the telecommunication industry.
It employs a (multi-modal) encoder-adapter-LLM decoder architecture and telecom-specific innovations.
GTS-LUM supports diverse time granularities, multi-modal data inputs, and aligning semantic information with behavior data.
Experiments validate the effectiveness of GTS-LUM, which outperforms LLM4Rec approaches in user behavior modeling in telecommunications.

Read Full Article

19 Likes

Arxiv

Image Credit: Arxiv

The Power of the Pareto Front: Balancing Uncertain Rewards for Adaptive Experimentation in scanning probe microscopy

Automated experimentation has the potential to revolutionize scientific discovery, but its effectiveness depends on well-defined optimization targets.
Multi-Objective Bayesian Optimization (MOBO) is applied to balance multiple, competing rewards in autonomous experimentation.
Using scanning probe microscopy (SPM) imaging, MOBO optimizes imaging parameters to enhance measurement quality, reproducibility, and efficiency.
MOBO offers a natural framework for human-in-the-loop decision-making, enabling researchers to fine-tune experimental trade-offs based on domain expertise.

Read Full Article

3 Likes

Arxiv

169

Image Credit: Arxiv

WaveHiTS: Wavelet-Enhanced Hierarchical Time Series Modeling for Wind Direction Nowcasting in Eastern Inner Mongolia

Wind direction forecasting plays a crucial role in optimizing wind energy production.
A novel model, WaveHiTS, has been developed to improve wind direction forecasting using wavelet transform and Neural Hierarchical Interpolation for Time Series (Hits).
Experiments conducted on real-world meteorological data show that WaveHiTS outperforms other deep learning models, transformer-based approaches, and hybrid models.
The proposed model achieves consistent accuracy for wind direction forecasting up to 60 minutes ahead, with significant improvements in RMSE values, vector correlation coefficients, and hit rates.

Read Full Article

10 Likes

Arxiv

173

Image Credit: Arxiv

Flexible Graph Similarity Computation With A Proactive Optimization Strategy

Graph Edit Distance (GED) is an important similarity measure in graph retrieval, allowing customizable operation costs.
A new approach called Graph Edit Network (GEN) is introduced to address the limitations of existing methods.
GEN incorporates operation costs in determining optimal graph mappings and proactively optimizes guidance from a graph perspective.
Results show that GEN outperforms state-of-the-art models in terms of error reduction and inference time reduction.

Read Full Article

10 Likes

Arxiv

Image Credit: Arxiv

TabKAN: Advancing Tabular Data Analysis using Kolmograv-Arnold Network

TabKAN is a novel framework that advances tabular data modeling using Kolmogorov-Arnold Networks (KANs).
KANs leverage learnable activation functions on edges, enhancing interpretability and training efficiency.
TabKAN demonstrates superior performance in supervised learning and outperforms classical and Transformer-based models in transfer learning scenarios.
KAN-based architectures bridge the gap between traditional machine learning and deep learning for structured data.

Read Full Article

2 Likes

Arxiv

116

Image Credit: Arxiv

NAPER: Fault Protection for Real-Time Resource-Constrained Deep Neural Networks

Fault tolerance in Deep Neural Networks (DNNs) deployed on resource-constrained systems presents unique challenges for high-accuracy applications with strict timing requirements.
The novel protection approach NAPER employs ensemble learning and heterogeneous model redundancy to achieve higher accuracy than traditional redundancy methods.
NAPER provides an efficient fault detection mechanism and a real-time scheduler to prioritize meeting deadlines and ensure uninterrupted operation during fault recovery.
Comparative evaluations show that NAPER offers 40% faster inference, 4.2% higher accuracy than TMR-based strategies, and effectively balances accuracy, reliability, and timeliness in real-time DNN applications.

Read Full Article

6 Likes

Arxiv

156

Image Credit: Arxiv

Disentangle and Regularize: Sign Language Production with Articulator-Based Disentanglement and Channel-Aware Regularization

Researchers propose a transformer-based sign language production (SLP) framework.
A pose autoencoder encodes sign poses into a compact latent space using an articulator-based disentanglement strategy.
A non-autoregressive transformer decoder predicts latent representations from sentence-level text embeddings.
Channel-aware regularization aligns predicted latent distributions with ground-truth encodings using KL-divergence loss.

Read Full Article

9 Likes

For uninterrupted reading, download the app