menu
techminis

A naukri.com initiative

google-web-stories
Home

>

ML News

ML News

source image

Arxiv

15h

read

354

img
dot

Image Credit: Arxiv

Handling Delay in Real-Time Reinforcement Learning

  • Real-time reinforcement learning (RL) involves challenges such as limited actions per second and observational delay.
  • Pipelining can address the limited actions issue, improving throughput and potential policy quality.
  • To tackle observational delay, a solution that leverages temporal skip connections and history-augmented observations is proposed.
  • Architectures with temporal skip connections achieve strong performance and parallel neuron computation can accelerate inference.

Read Full Article

like

21 Likes

source image

Arxiv

15h

read

117

img
dot

Image Credit: Arxiv

A Survey on Unlearnable Data

  • Unlearnable data (ULD) is a defense technique to prevent machine learning models from learning meaningful patterns from specific data, protecting data privacy and security.
  • This survey provides a comprehensive review of ULD, including data generation methods, evaluation metrics, and practical applications.
  • Different ULD approaches are compared and contrasted in terms of unlearnability, imperceptibility, efficiency, and robustness.
  • The survey also explores challenges and future research directions to enhance the effectiveness and applicability of ULD in data protection for machine learning.

Read Full Article

like

7 Likes

source image

Arxiv

15h

read

120

img
dot

Image Credit: Arxiv

Redundant feature screening method for human activity recognition based on attention purification mechanism

  • Researchers propose a universal attention feature purification mechanism (MSAP) for Human Activity Recognition (HAR).
  • The mechanism effectively reduces feature redundancy caused by multi-scale features in wearable devices.
  • A network correction module is designed to mitigate inherent problems in deep networks.
  • Experiments show that the proposed method model effectively reduces redundant features with little resource consumption.

Read Full Article

like

7 Likes

source image

Arxiv

15h

read

73

img
dot

Image Credit: Arxiv

Bridging conformal prediction and scenario optimization

  • Conformal prediction and scenario optimization are two statistical learning frameworks used to certify decisions made using data.
  • While these frameworks have been extensively studied and yield similar results, no clear connection between them has been established.
  • This research focuses on vanilla conformal prediction and demonstrates how to choose appropriate score functions and set predictor maps to recover bounds on the probability of constraint violation associated with scenario programs.
  • The study also establishes a theoretical bridge between conformal prediction and scenario optimization, allowing for analysis of calibration conditional conformal prediction.

Read Full Article

like

4 Likes

source image

Arxiv

15h

read

361

img
dot

Image Credit: Arxiv

Partial Transportability for Domain Generalization

  • A new research paper introduces results for bounding the value of a functional of the target distribution, such as the generalization error of a classifier, given data from source domains and assumptions about the data generating mechanisms.
  • The paper builds on the theory of partial identification and transportability to provide the first general estimation technique for transportability problems.
  • The authors adapt existing parameterization schemes, such as Neural Causal Models, to encode the necessary structural constraints for cross-population inference.
  • The paper also proposes a gradient-based optimization scheme to make scalable inferences in practice, and the results are supported by experiments.

Read Full Article

like

21 Likes

source image

Arxiv

15h

read

7

img
dot

Image Credit: Arxiv

Autonomous Learning with High-Dimensional Computing Architecture Similar to von Neumann's

  • Researchers have developed a computing architecture for autonomous learning that resembles traditional (von Neumann) computing with numbers but performs operations on high-dimensional vectors.
  • The architecture includes a high-capacity memory for vectors, similar to random-access memory (RAM) for numbers, and is inspired by models of human and animal learning.
  • This approach, which aligns with ideas from psychology, biology, and traditional computing, provides insights into how brains compute and can potentially be applied to enable learning by robots.
  • To realize the vision of computing with minimal material and energy usage, further development of a mathematical theory and large-scale experiments are required.

Read Full Article

like

Like

source image

Arxiv

15h

read

117

img
dot

Image Credit: Arxiv

Graph-Eq: Discovering Mathematical Equations using Graph Generative Models

  • The ability to discover meaningful, accurate, and concise mathematical equations that describe datasets is valuable across various domains.
  • Equations offer explicit relationships between variables, enabling deeper insights into underlying data patterns.
  • In this work, a deep graph generative model called Graph-EQ is proposed for efficient equation discovery.
  • Graph-EQ uses a conditional variational autoencoder (CVAE) to learn a rich latent representation of the equation space and has shown success in discovering ground-truth equations for given datasets.

Read Full Article

like

7 Likes

source image

Arxiv

15h

read

208

img
dot

Image Credit: Arxiv

Simple Feedfoward Neural Networks are Almost All You Need for Time Series Forecasting

  • Simple feedforward neural networks (SFNNs) can achieve performance on par with, or even exceeding, advanced models like Transformers and graph neural networks (GNNs) in time series forecasting.
  • SFNNs are simpler, smaller, faster, and more robust compared to the state-of-the-art models.
  • Even in cases where modeling interactions between multiple series is needed, a basic multivariate SFNN can still deliver competitive results.
  • SFNNs serve as a strong baseline and future time series forecasting methods should be compared against them.

Read Full Article

like

12 Likes

source image

Arxiv

15h

read

179

img
dot

Image Credit: Arxiv

A Survey of Reinforcement Learning-Based Motion Planning for Autonomous Driving: Lessons Learned from a Driving Task Perspective

  • Reinforcement learning (RL) has emerged as a promising approach for motion planning challenges in autonomous driving (AD).
  • This survey provides a comprehensive review of RL-based motion planning, focusing on lessons learned from a driving task perspective.
  • It outlines the fundamentals of RL methodologies and analyzes their applications in motion planning, considering scenario-specific features and task requirements.
  • The survey also identifies frontier challenges and proposes strategies for overcoming unresolved issues in RL-based motion planning.

Read Full Article

like

10 Likes

source image

Arxiv

15h

read

212

img
dot

Image Credit: Arxiv

Dynamic Operating System Scheduling Using Double DQN: A Reinforcement Learning Approach to Task Optimization

  • An operating system scheduling algorithm based on Double DQN (Double Deep Q network) is proposed.
  • The algorithm dynamically adjusts task priority and resource allocation strategy, improving task completion efficiency, system throughput, and response speed.
  • Experimental results show that the Double DQN algorithm performs well under different system loads and is particularly effective for I/O intensive tasks.
  • The algorithm also demonstrates high optimization ability in resource utilization and can intelligently adjust resource allocation based on the system state.

Read Full Article

like

12 Likes

source image

Arxiv

15h

read

175

img
dot

Image Credit: Arxiv

Data-Driven Forecasting of High-Dimensional Transient and Stationary Processes via Space-Time Projection

  • Space-Time Projection (STP) is introduced as a data-driven forecasting approach for high-dimensional and time-resolved data.
  • STP computes extended space-time proper orthogonal modes from training data to generate forecasts by projecting these modes onto new data.
  • The method relies on the orthogonality and optimal correlation of the modes, and no additional hyperparameters are required.
  • Comparative studies with LSTM neural networks showed that STP consistently provided more accurate forecasts.

Read Full Article

like

10 Likes

source image

Arxiv

15h

read

292

img
dot

Image Credit: Arxiv

A Low-complexity Structured Neural Network to Realize States of Dynamical Systems

  • This paper introduces a structured neural network (StNN) for realizing states of dynamical systems by leveraging data-driven learning.
  • The StNN utilizes a low-complexity operator called the Hankel operator, derived from time-delay measurements, to solve dynamical systems.
  • Numerical simulations comparing the StNN with other techniques show that it reduces the number of parameters and computational complexity.
  • The proposed StNN enables the prediction and understanding of future states in state-space dynamical systems.

Read Full Article

like

17 Likes

source image

Arxiv

15h

read

219

img
dot

Image Credit: Arxiv

Steering Large Agent Populations using Mean-Field Schrodinger Bridges with Gaussian Mixture Models

  • The Mean-Field Schrodinger Bridge (MFSB) problem aims to find the minimum effort control policy for a swarm of cooperative agents.
  • New efficient parameterization is proposed to approximate MFSB solutions for Gaussian Mixture Model boundary distributions.
  • The proposed approach uses a mixture of elementary policies to solve a Gaussian-to-Gaussian Covariance Steering problem.
  • The method can handle probabilistic hard constraints and is applied to various numerical examples.

Read Full Article

like

13 Likes

source image

Arxiv

15h

read

65

img
dot

Image Credit: Arxiv

Unimodal-driven Distillation in Multimodal Emotion Recognition with Dynamic Fusion

  • Unimodal-driven Distillation in Multimodal Emotion Recognition with Dynamic Fusion
  • SUMMER is a novel heterogeneous multimodal integration framework for emotion recognition
  • It leverages Mixture of Experts with Hierarchical Cross-modal Fusion and Interactive Knowledge Distillation
  • Experiments show that SUMMER outperforms state-of-the-art methods in recognizing minority and semantically similar emotions

Read Full Article

like

3 Likes

source image

Arxiv

15h

read

84

img
dot

Image Credit: Arxiv

PDSL: Privacy-Preserved Decentralized Stochastic Learning with Heterogeneous Data Distribution

  • In the paradigm of decentralized learning, a group of agents collaborates to learn a global model using distributed datasets without a central server.
  • The heterogeneity of the local data across agents makes learning a robust global model challenging.
  • PDSL is a privacy-preserved decentralized stochastic learning algorithm that addresses these challenges using Shapley values to measure neighbor contributions and differential privacy to prevent privacy leakage.
  • The PDSL algorithm demonstrates efficacy in privacy preservation and convergence, supported by theoretical analysis and extensive experiments.

Read Full Article

like

5 Likes

For uninterrupted reading, download the app