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

A naukri.com initiative

New

Home

ML News

Arxiv

139

Image Credit: Arxiv

AMAD: AutoMasked Attention for Unsupervised Multivariate Time Series Anomaly Detection

Unsupervised multivariate time series anomaly detection (UMTSAD) is important in various domains.
Deep learning models based on the Transformer and self-attention mechanisms have shown impressive results in UMTSAD.
However, these models have limitations in generalizing to diverse anomaly situations without labeled data.
To address this, the proposed model, AMAD, integrates AutoMasked Attention for UMTSAD scenarios, providing a robust and adaptable solution.

Read Full Article

8 Likes

Arxiv

165

Image Credit: Arxiv

GRAIN: Multi-Granular and Implicit Information Aggregation Graph Neural Network for Heterophilous Graphs

Graph neural networks (GNNs) have shown significant success in learning graph representations.
However, GNNs often fail to outperform simple MLPs on heterophilous graph tasks.
To overcome this, the Granular and Implicit Graph Network (GRAIN) is proposed, a novel GNN model specifically designed for heterophilous graphs.
GRAIN effectively integrates local and global information, resulting in improved node representations.

Read Full Article

9 Likes

Arxiv

291

Image Credit: Arxiv

A Neuro-inspired Interpretation of Unlearning in Large Language Models through Sample-level Unlearning Difficulty

Driven by privacy protection laws and regulations, unlearning in Large Language Models (LLMs) is gaining attention.
Researchers propose a metric called Memory Removal Difficulty (MRD) to quantify sample-level unlearning difficulty.
The study explores the characteristics of hard-to-unlearn and easy-to-unlearn samples in LLM unlearning.
An MRD-based weighted sampling method is proposed to optimize existing unlearning algorithms, improving efficiency and effectiveness.

Read Full Article

17 Likes

Arxiv

295

Image Credit: Arxiv

Bridging the Gap Between Preference Alignment and Machine Unlearning

Mainstream methods like Reinforcement Learning with Human Feedback (RLHF) face challenges in Preference Alignment (PA) for Large Language Models (LLMs).
High-quality datasets of positive preference examples are costly and computationally intensive, limiting their use in low-resource scenarios.
LLM unlearning technique presents a promising alternative by directly removing the influence of negative examples.
A framework called Unlearning to Align (U2A) is proposed to optimize the selection and unlearning of negative examples for improved PA performance.

Read Full Article

17 Likes

Discover more

Arxiv

388

Image Credit: Arxiv

Robust and Noise-resilient Long-Term Prediction of Spatiotemporal Data Using Variational Mode Graph Neural Networks with 3D Attention

This paper presents a method to improve the robustness of spatiotemporal long-term prediction using variational mode graph convolutional networks (VMGCN) with 3D channel attention.
The method incorporates i.i.d. Gaussian noise to a large traffic volume dataset and applies variational mode decomposition to model the corrupted signal.
A 3D attention mechanism is integrated to learn spatial, temporal, and channel correlations and to suppress noise while highlighting significant modes in the spatiotemporal signals.
The proposed method outperforms baseline models in terms of long-term prediction accuracy, robustness to noise, and improved performance with mode truncation.

Read Full Article

23 Likes

Arxiv

240

Image Credit: Arxiv

Hyperparameter Optimisation with Practical Interpretability and Explanation Methods in Probabilistic Curriculum Learning

Hyperparameter optimisation is crucial for achieving strong performance in reinforcement learning (RL).
Probabilistic Curriculum Learning (PCL) is a curriculum learning strategy designed to improve RL performance.
This paper provides an empirical analysis of hyperparameter interactions and their effects on the performance of a PCL algorithm.
The study presents strategies to refine hyperparameter search spaces and introduces a novel SHAP-based interpretability approach for analyzing hyperparameter impacts in RL.

Read Full Article

14 Likes

Arxiv

397

Image Credit: Arxiv

Benchmarking Convolutional Neural Network and Graph Neural Network based Surrogate Models on a Real-World Car External Aerodynamics Dataset

A comparative evaluation of surrogate modeling approaches for predicting drag on a real-world car aerodynamics dataset was conducted.
The evaluation compared a Convolutional Neural Network (CNN) model using a signed distance field as input and a commercial tool based on Graph Neural Networks (GNN) processing a surface mesh.
The CNN-based method achieved a mean absolute error of 2.3 drag counts, while the GNN-based method achieved 3.8.
Both methods achieved approximately 77% accuracy in predicting the direction of drag change relative to the baseline geometry.

Read Full Article

23 Likes

Arxiv

414

Image Credit: Arxiv

Clustering and novel class recognition: evaluating bioacoustic deep learning feature extractors

A paper in arXiv discusses the evaluation of bioacoustic deep learning feature extractors for clustering and novel class recognition.
The research aims to address the limitation of benchmarking classification scores, which is specific to the training data and does not allow comparison across different taxonomic groups.
The study analyzes the embeddings generated by 15 bioacoustic models to evaluate their adaptability and generalization potential.
Clustering and kNN classification are used to structure the embedding spaces, allowing comparison of feature extractors independent of their classifiers.

Read Full Article

24 Likes

Arxiv

131

Image Credit: Arxiv

Plastic tensor networks for interpretable generative modeling

A structural optimization scheme for a single-layer nonnegative adaptive tensor tree (NATT) that models a target probability distribution is proposed.
The NATT scheme is interpretable as a probabilistic graphical model and is compared to the Born machine adaptive tensor tree (BMATT) optimization scheme on generative modeling tasks.
The results show that the single-layer NATT scheme has comparable model performance to the BMATT scheme in terms of minimizing the negative log-likelihood.
The study explores tasks such as deducing the structure of binary bitwise operations, learning the internal structure of random Bayesian networks, and constructing a cladogram from mitochondrial DNA sequences.

Read Full Article

7 Likes

Arxiv

254

Image Credit: Arxiv

PETNet -- Coincident Particle Event Detection using Spiking Neural Networks

Spiking neural networks (SNN) are investigated for detecting photon coincidences in positron emission tomography (PET) data.
PETNet interprets detector hits as a binary-valued spike train and learns to identify photon coincidence pairs.
PETNet outperforms the state-of-the-art classical algorithm with a maximal coincidence detection F1 of 95.2%.
PETNet predicts photon coincidences up to 36 times faster than the classical approach, demonstrating the potential of SNNs in particle physics applications.

Read Full Article

15 Likes

Arxiv

418

Image Credit: Arxiv

Efficient Deployment of Spiking Neural Networks on SpiNNaker2 for DVS Gesture Recognition Using Neuromorphic Intermediate Representation

Spiking Neural Networks (SNNs) are highly energy-efficient during inference, making them suitable for deployment on neuromorphic hardware.
A benchmark study evaluates two quantization pipelines for fixed-point computations in SNNs optimized for the SpiNNaker2 chip.
The first approach employs post training quantization (PTQ) with percentile-based threshold scaling.
The second method uses quantization aware training (QAT) with adaptive threshold scaling, both achieving accurate 8-bit on-chip inference.

Read Full Article

25 Likes

Arxiv

257

Image Credit: Arxiv

FedMerge: Federated Personalization via Model Merging

One global model in federated learning (FL) might not be sufficient to serve many clients with non-IID tasks and distributions.
The paper proposes a novel approach called FedMerge that can create a personalized model per client by merging multiple global models with optimized weights.
FedMerge allows a few global models to serve many non-IID clients without requiring further local fine-tuning.
The approach outperforms existing FL methods across different non-IID settings in terms of performance.

Read Full Article

15 Likes

Arxiv

198

Image Credit: Arxiv

Beware of "Explanations" of AI

Understanding the decisions made and actions taken by AI systems is a challenge.
Explainable Artificial Intelligence (XAI) aims to provide explanations to enhance trust and adoption.
Defining what constitutes a 'good' explanation is difficult due to various factors.
Poorly designed explanations can lead to risks and harm, including wrong decisions and privacy violations.

Read Full Article

11 Likes

Arxiv

139

Image Credit: Arxiv

Robust Classification with Noisy Labels Based on Posterior Maximization

This paper investigates the robustness of an f-divergence-based class of objective functions, referred to as f-PML, to label noise in supervised classification.
The study shows that, in the presence of label noise, the f-PML objective functions can be corrected to obtain a neural network that matches the clean dataset.
An alternative correction approach is proposed to refine the posterior estimation during the test phase for neural networks trained with label noise.
The paper demonstrates that the f-PML objective functions are robust to symmetric label noise and can achieve competitive performance with refined training strategies.

Read Full Article

8 Likes

Arxiv

380

Image Credit: Arxiv

Regret Bounds for Robust Online Decision Making

A framework is proposed for robust online decision making, allowing multivalued models.
The framework introduces convex sets of probability distributions for decision outcomes.
Nature can choose distributions from the set in an arbitrary and non-oblivious manner.
The framework demonstrates improved regret bounds in robust linear bandits and tabular robust online reinforcement learning.

Read Full Article

22 Likes

For uninterrupted reading, download the app