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

A naukri.com initiative

New

Home

ML News

Arxiv

16h

227

Image Credit: Arxiv

Gradient-Based Neuroplastic Adaptation for Concurrent Optimization of Neuro-Fuzzy Networks

A new approach, gradient-based neuroplastic adaptation, is proposed for optimizing Neuro-fuzzy networks (NFNs) parameters and structure concurrently.
NFNs are symbolic function approximations with advantages like transparency and universal function approximation ability.
The traditional sequential design process for NFNs is inefficient, leading to suboptimal architecture; the new approach addresses this limitation.
Empirical evidence shows the effectiveness of the new method in training NFNs with online reinforcement learning to excel in vision-based video game scenarios.

Read Full Article

13 Likes

Arxiv

16h

Image Credit: Arxiv

Multi-task parallelism for robust pre-training of graph foundation models on multi-source, multi-fidelity atomistic modeling data

Graph foundation models using graph neural networks are being used for atomistic modeling to handle multi-source, multi-fidelity data during pre-training.
Recent studies employ multi-task learning where shared layers process atomistic structures initially regardless of the source, routing them to different decoding heads for data-specific predictions.
A new multi-task parallelism method is proposed to distribute each head across computing resources with GPU acceleration, implemented in the open-source HydraGNN architecture.
The method was trained on over 24 million structures from five datasets and tested on supercomputers like Perlmutter, Aurora, and Frontier, showing efficient scaling on heterogeneous computing architectures.

Read Full Article

5 Likes

Arxiv

16h

345

Image Credit: Arxiv

Why Neural Network Can Discover Symbolic Structures with Gradient-based Training: An Algebraic and Geometric Foundation for Neurosymbolic Reasoning

Researchers have developed a theoretical framework explaining how neural networks can naturally discover discrete symbolic structures through gradient-based training.
By lifting neural parameters to a measure space and utilizing Wasserstein gradient flow, the framework demonstrates the emergence of symbolic phenomena under geometric constraints like group invariance.
The framework highlights the decoupling of gradient flow into independent optimization trajectories based on potential functions and a reduction in degrees of freedom, leading to the encoding of algebraic constraints relevant to the task.
The research establishes data scaling laws connecting representational capacity to group invariance, enabling neural networks to transition from high-dimensional exploration to compositional representations aligned with algebraic operations, offering insights for designing neurosymbolic systems.

Read Full Article

20 Likes

Arxiv

16h

Image Credit: Arxiv

The Cost of Avoiding Backpropagation

Forward-mode automatic differentiation (FmAD) and zero-order (ZO) optimization have been proposed as memory-efficient alternatives to backpropagation (BP) for gradient computation.
A new study presents a comparison of BP, FmAD, and ZO methods, highlighting theoretical and empirical findings.
Theoretical analysis suggests that FmAD and ZO reduce memory usage but at the cost of accuracy, convergence speed, and computation compared to BP with checkpointing.
Empirical experiments on large models demonstrate that BP with checkpointing outperforms FmAD and ZO variants, indicating BP with checkpointing as the most effective strategy for model training in memory-constrained settings.

Read Full Article

5 Likes

Discover more

Arxiv

16h

334

Image Credit: Arxiv

Koopman operator-based discussion on partial observation in stochastic systems

The discussion revolves around the effects of partial observation in stochastic systems using the Koopman operator theory.
Data-driven algorithms based on the Koopman operator theory have shown progress in handling partial observations, aiming to connect with the Mori-Zwanzig formalism.
The importance of differentiating the state space and function space in stochastic systems is emphasized in the analysis.
Numerical experiments demonstrate the benefits of the delay embedding technique for partial observation in stochastic systems, revealing a power-law behavior in the accuracy of the additive noise amplitude.

Read Full Article

20 Likes

Arxiv

16h

Image Credit: Arxiv

A Survey of Continual Reinforcement Learning

Reinforcement Learning (RL) has seen progress in solving decision-making problems, leveraging deep neural networks.
Continual Reinforcement Learning (CRL) has emerged to address RL limitations by enabling continuous learning and adaptation to new tasks.
A survey on CRL covers core concepts, challenges, and methodologies, reviewing existing works and proposing a new taxonomy of CRL methods.
The analysis underscores the unique challenges of CRL and offers insights for future research directions in this field.

Read Full Article

Arxiv

16h

213

Image Credit: Arxiv

TOAST: Task-Oriented Adaptive Semantic Transmission over Dynamic Wireless Environments

TOAST (Task-Oriented Adaptive Semantic Transmission) is a framework introduced to tackle multi-task optimization in dynamic wireless environments for 6G networks.
The framework includes components like adaptive task balancing using deep reinforcement learning, Low-Rank Adaptation (LoRA) mechanisms, and an Elucidating diffusion model for feature restoration.
Experiments show that TOAST outperforms baseline approaches in improving classification accuracy and reconstruction quality, especially in low Signal-to-Noise Ratio (SNR) conditions.
The framework aims to shift communication from bit-centric to semantic-aware, emphasizing task-relevant information in wireless transmissions.

Read Full Article

12 Likes

Arxiv

16h

124

Image Credit: Arxiv

GuiderNet: A Meta-Learning Framework for Optimizing Quantum Circuit Geometry and Mitigating Barren Plateaus

GuiderNet is a meta-learning framework designed to optimize quantum circuit geometry and address issues with barren plateaus in Variational Quantum Algorithms.
It conditions Parameterized Quantum Circuits (PQCs) using data-dependent parameter shifts to minimize the log condition number of the Fubini-Study metric tensor.
GuiderNet has shown significant improvements in tasks like the Kaggle Diabetes classification by reducing training loss, increasing test accuracy, and improving generalization in quantum machine learning.
The framework suppresses gradient explosion, stabilizes parameter updates, and enhances trainability, demonstrating its potential to mitigate barren plateaus and ill-conditioning in quantum algorithms.

Read Full Article

7 Likes

Arxiv

16h

Image Credit: Arxiv

Physics-informed network paradigm with data generation and background noise removal for diverse distributed acoustic sensing applications

A new physics-informed DAS neural network paradigm is proposed for diverse distributed acoustic sensing applications.
This paradigm does not require real-world events data for training, as it generates DAS events data through physical modeling.
The network is trained to remove background noise in DAS data, showing effectiveness in event identification and fault monitoring applications.
The paradigm demonstrates generalization in different sites and achieves a fault diagnosis accuracy of 91.8% in belt conveyor field without test site data for training.

Read Full Article

Arxiv

16h

356

Image Credit: Arxiv

Optimal Return-to-Go Guided Decision Transformer for Auto-Bidding in Advertisement

Advertisers in online advertising use auto-bidding tools in ad auctions on demand-side platforms.
Researchers have introduced the R* Decision Transformer (R* DT) to enhance automated bidding systems by addressing challenges in conventional Decision Transformers (DT), such as lack of preset return-to-go (RTG) values and mixed-quality training data.
The R* DT is developed in three steps: R DT stores actions based on state and RTG, R^ DT predicts the highest RTG for a state to derive a suboptimal policy, and R* DT generates trajectories based on R^ DT to improve training data quality and move towards an optimal policy.
Tests on a public bidding dataset demonstrate the effectiveness of R* DT in handling mixed-quality trajectories and improving the RTG of bidding actions.

Read Full Article

21 Likes

Arxiv

16h

264

Image Credit: Arxiv

Binned semiparametric Bayesian networks

This paper introduces a new type of probabilistic semiparametric model that utilizes data binning to improve computational efficiency in nonparametric distributions.
Two new conditional probability distributions, sparse binned kernel density estimation and Fourier kernel density estimation, are developed for the binned semiparametric Bayesian networks.
The models address the curse of dimensionality by employing sparse tensors and limiting the number of parent nodes in conditional probability calculations.
Experiments with synthetic data and datasets from the UCI Machine Learning repository show that the binned semiparametric Bayesian networks achieve similar performance to non-binned models in terms of structural learning and log-likelihood estimations, but with significantly higher speed.

Read Full Article

15 Likes

Arxiv

16h

Image Credit: Arxiv

GKNet: Graph Kalman Filtering and Model Inference via Model-based Deep Learning

Researchers have introduced GKNet, a graph-aware state space model for inference tasks with time series over graphs.
The model includes a graph-induced state equation driven by noise over graph edges and a graph-filtered observation model.
Parameters in both state and observation models are learned from partially observed data for prediction and imputation.
To enhance scalability, a deep learning architecture inspired by Kalman neural networks is employed for end-to-end learning and parameter tracking.

Read Full Article

2 Likes

Arxiv

16h

Image Credit: Arxiv

TROFI: Trajectory-Ranked Offline Inverse Reinforcement Learning

TROFI is a new approach in offline reinforcement learning that aims to train agents without a predefined reward function.
It first learns a reward function from human preferences to label the dataset, enabling training of the policy.
Experiments on the D4RL benchmark show that TROFI outperforms baselines and performs similarly to using the ground truth reward.
The efficacy of TROFI is validated in a 3D game environment, emphasizing the importance of a well-engineered reward function in reinforcement learning.

Read Full Article

1 Like

Arxiv

16h

Image Credit: Arxiv

Hyper-modal Imputation Diffusion Embedding with Dual-Distillation for Federated Multimodal Knowledge Graph Completion

Researchers propose a model for Federated Multimodal Knowledge Graph Completion (FedMKGC) to predict missing links in decentralized knowledge graphs without sharing sensitive information.
The proposed framework, MMFeD3-HidE, addresses challenges such as incomplete entity embeddings and client heterogeneity in FedMKGC.
MMFeD3-HidE consists of a Hyper-modal Imputation Diffusion Embedding model (HidE) for recovering multimodal distributions and Multimodal Federated Dual Distillation (MMFeD3) for transferring knowledge between clients and the server.
Experiments on the proposed benchmark demonstrate the effectiveness, semantic consistency, and convergence robustness of MMFeD3-HidE for Federated Multimodal Knowledge Graph Completion.

Read Full Article

3 Likes

Arxiv

16h

Image Credit: Arxiv

UniCA: Adapting Time Series Foundation Model to General Covariate-Aware Forecasting

Unified Covariate Adaptation (UniCA) bridges Time Series Foundation Models (TSFMs) with general covariate-aware forecasting to handle diverse covariates like categorical variables and multimodal data.
UniCA performs covariate homogenization to transform heterogeneous covariates into homogeneous series representations and fuses them using an attention-based fusion mechanism.
Experiments on various forecasting benchmarks demonstrate the superiority of UniCA in incorporating extra covariate information while preserving the generalization ability of TSFMs.
Code related to UniCA is available on GitHub for further exploration.

Read Full Article

3 Likes

For uninterrupted reading, download the app