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

A naukri.com initiative

New

Home

ML News

Arxiv

412

Image Credit: Arxiv

ClimateBench-M: A Multi-Modal Climate Data Benchmark with a Simple Generative Method

Climate science studies the structure and dynamics of Earth's climate system and seeks to understand how climate changes over time.
A multi-modal climate benchmark called ClimateBench-M has been developed by aligning time series climate data, extreme weather events data, and satellite image data.
The benchmark includes a simple but strong generative method that delivers competitive performance in weather forecasting, thunderstorm alerts, and crop segmentation tasks.
The data and code of ClimateBench-M are publicly available on GitHub.

Read Full Article

24 Likes

Arxiv

Image Credit: Arxiv

MicroNAS: An Automated Framework for Developing a Fall Detection System

MicroNAS is an automated neural architecture search tool designed for developing models optimized for microcontrollers with small memory resources.
It uses a novel method that considers the memory size of the target microcontroller to optimize convolutional neural network and gated recurrent unit architectures.
A comparison is made between memory-driven model optimization and traditional pruning methods, demonstrating the effectiveness of MicroNAS.
MicroNAS achieved higher F1-scores in developing a fall detection system (FDS), showing its potential in real-time FDS development for microcontroller platforms with limited memory.

Read Full Article

Arxiv

327

Image Credit: Arxiv

Multi-Selection for Recommendation Systems

Researchers have developed a multi-selection model for answering differentially private queries in recommendation systems.
The model allows the server to send multiple recommendations and a 'local model' to the user.
Users can then use the local model to select the item that best matches their private features.
The multi-selection paradigm achieves an average recommendation utility of approximately 97% while maintaining local differential privacy.

Read Full Article

19 Likes

Arxiv

315

Image Credit: Arxiv

Unifying and extending Diffusion Models through PDEs for solving Inverse Problems

Diffusion models have been used to solve probabilistic inverse problems in computer vision and scientific machine learning.
This study introduces a new approach to derive diffusion models using ideas from linear partial differential equations.
The new approach enables a unified derivation of multiple formulations and sampling strategies.
The study also explores the applications of conditional diffusion models in solving density estimation problems and inverse problems.

Read Full Article

18 Likes

Discover more

Arxiv

Image Credit: Arxiv

LoRI: Reducing Cross-Task Interference in Multi-Task Low-Rank Adaptation

Low-Rank Adaptation (LoRA) has emerged as a popular parameter-efficient fine-tuning (PEFT) method for Large Language Models (LLMs), yet it still incurs notable overhead and suffers from parameter interference in multi-task scenarios.
LoRI with Reduced Interference (LoRI) is a simple yet effective approach that reduces the number of trainable parameters while maintaining strong task performance.
LoRI leverages orthogonality between adapter subspaces to minimize cross-task interference in adapter merging, and uses sparsity to mitigate catastrophic forgetting for continual learning.
Experiments across various tasks show that LoRI outperforms full fine-tuning and existing PEFT methods, while using significantly fewer trainable parameters.

Read Full Article

1 Like

Arxiv

279

Image Credit: Arxiv

Multi-Modal Data Fusion for Moisture Content Prediction in Apple Drying

Fruit drying is an important process in food manufacturing for reducing moisture content, ensuring product safety, and extending shelf life.
A new multi-modal data fusion framework has been proposed to improve the accuracy of moisture content prediction in apple drying.
The framework effectively combines tabular data (process parameters) and high-dimensional image data (images of dried apple slices).
Experimental validation demonstrated significant improvement in predictive accuracy compared to existing methods.

Read Full Article

16 Likes

Arxiv

391

Image Credit: Arxiv

Traversal Learning Coordination For Lossless And Efficient Distributed Learning

Researchers introduce Traversal Learning (TL), a novel approach to address the decreased quality in distributed learning paradigms such as Federated Learning (FL), Split Learning (SL), and SplitFed Learning (SFL). TL adopts a unique strategy where the model traverses nodes during forward propagation and performs backward propagation on the orchestrator, effectively implementing centralized learning (CL) principles within a distributed environment. TL outperformed other DL methods and improved accuracy across various datasets representing different domains. TL represents a significant advancement in DL methodologies that preserves data privacy while maintaining performance.

Read Full Article

23 Likes

Arxiv

243

Image Credit: Arxiv

Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving

Apt-Serve is a scalable framework designed to enhance effective throughput in large language model (LLM) inference serving systems.
It addresses the bottleneck caused by a memory-intensive KV cache and rigid batch composition in existing systems.
Apt-Serve combines KV cache with a memory-efficient hidden cache for reusable input hidden state vectors, allowing larger batch sizes and improved request concurrency.
Evaluations show that Apt-Serve achieves up to 8.8x improvement in effective throughput compared to state-of-the-art inference serving systems.

Read Full Article

14 Likes

Arxiv

294

Image Credit: Arxiv

GPT Carry-On: Training Foundation Model for Customization Could Be Simple, Scalable and Affordable

Researchers propose a framework for customizing large language foundation models (LLMs) for specific users or tasks.
The framework involves training an additional branch of transformer blocks on the final-layer embedding of pretrained LLMs, and using a carry-on module to merge the base models.
Multiple layers or LLMs specialized in different domains can be combined to create a customized LLM for a new task.
The proposed approach allows outsourcing most computation of the training job on inference nodes, reducing the memory and computation requirements.

Read Full Article

17 Likes

Arxiv

Image Credit: Arxiv

Adversarial Subspace Generation for Outlier Detection in High-Dimensional Data

Outlier detection in high-dimensional tabular data is challenging due to the Multiple Views effect, where data is distributed across multiple lower-dimensional subspaces.
A new theoretical framework called Myopic Subspace Theory (MST) is introduced, which mathematically formulates the Multiple Views effect and writes subspace selection as a stochastic optimization problem.
V-GAN, a generative method trained based on MST, is presented to avoid exhaustive search over the feature space while preserving the intrinsic data structure.
Experiments on real-world datasets demonstrate that using V-GAN subspaces leads to improved one-class classification performance compared to existing methods, confirming the theoretical guarantees and practical viability of the approach.

Read Full Article

Arxiv

252

Image Credit: Arxiv

PoGO: A Scalable Proof of Useful Work via Quantized Gradient Descent and Merkle Proofs

Researchers propose a design called Proof of Gradient Optimization (PoGO) for blockchain consensus.
PoGO involves miners producing verifiable evidence of training large-scale machine-learning models.
The design incorporates quantized gradients to reduce storage and computation requirements while enabling verification of real progress in lowering the model's loss.
The system employs Merkle proofs over the full 32-bit model and allows verifiers to issue positive or negative attestations.

Read Full Article

15 Likes

Arxiv

121

Image Credit: Arxiv

Conditional Conformal Risk Adaptation

Uncertainty quantification is crucial in image segmentation, particularly for high-stakes applications like medical imaging.
Conformal Risk Adaptation (CRA) is introduced to improve conditional risk control for segmentation tasks.
A novel framework connects conformal risk control with conformal prediction, enhancing any score function.
Calibrated Conformal Risk Adaptation (CCRA) and CCRA-S further improve conditional risk control for segmentation.

Read Full Article

7 Likes

Arxiv

268

Image Credit: Arxiv

Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather

Predictions on subseasonal-to-seasonal (S2S) timescales remain challenging due to chaos in the climate system.
Deep learning architectures have been developed to improve S2S forecast skill by incorporating teleconnections such as the stratospheric polar vortex and Madden-Julian Oscillation.
The ViT-LSTM model outperforms other AI models and ECMWF's hindcasts in predicting North Atlantic-European weather regimes, specifically Scandinavian Blocking and Atlantic Ridge patterns.
The study demonstrates the potential of deep learning methods in enhancing subseasonal forecasting and providing new insights into atmospheric dynamics and predictability.

Read Full Article

16 Likes

Arxiv

411

Image Credit: Arxiv

Predicting the Lifespan of Industrial Printheads with Survival Analysis

Accurately predicting the lifespan of critical device components is essential for maintenance planning and production optimization.
In this work, survival analysis is used to predict the lifespan of production printheads developed by Canon Production Printing.
Five techniques, including the Kaplan-Meier estimator, Cox proportional hazard model, Weibull accelerated failure time model, random survival forest, and gradient boosting, are applied to estimate survival probabilities and failure rates.
Quantitative evaluation demonstrates that survival analysis outperforms industry-standard baseline methods for printhead lifespan prediction.

Read Full Article

24 Likes

Arxiv

109

Image Credit: Arxiv

Prediction of Usage Probabilities of Shopping-Mall Corridors Using Heterogeneous Graph Neural Networks

Researchers present a method based on graph neural networks (GNN) to predict usage probabilities of shopping-mall corridors.
The method utilizes heterogeneous graph networks created from floorplans of malls, including corridors, shops, and entrances.
Features such as shop area and usage categories are used for prediction, along with connections represented as corridor paths.
The method incorporates a supervised deep-learning workflow, including latent feature representation learning and multilayer perceptrons (MLP) for final predictions.

Read Full Article

6 Likes

For uninterrupted reading, download the app