Convolutional Neural Networks Training for Autonomous Robotics

Alexander Lozhkin,Konstantin Maiorov,Pavol Bozek

doi:10.2478/mspe-2021-0010

Alexander Lozhkin, Konstantin Maiorov + Show 1 more

Open Access

https://doi.org/10.2478/mspe-2021-0010

Copy DOI

Abstract

Abstract The article discusses methods for accelerating the operation of convolutional neural networks for autonomous robotics learning. The analysis of the theoretical possibility of modifying the neural network learning mechanism is carried out. Classic semiotic analysis and the theory of neural networks is proposed to union. An assumption is made about the possibility of using the symmetry mechanism to accelerate the training of convolutional neural networks. A multilayer neural network to represent how space is an attempt has been made. The conclusion was based on the laws on the plane obtained earlier. The derivation of formulas turned out to be impossible due to the problems of modern mathematics. A new approach is proposed, which involves combining the gradient descent algorithm and the stochastic completion of convolutional filters by the principles of symmetries. The identified algorithms allow increasing the learning rate from 5% to 15%, depending on the problem that the neural network solves.

Highlights

Modern autonomous robots cannot be imagined without artificial intelligence systems, namely neural networks
Simple modification for neural systems The idea behind a new approach to speed up the training of convolutional networks is to use this symmetry mechanism for faster network convergence
The approach consists of two stages: symmetric initialization of convolutional filters and stochastic symmetric filtering

Summary

Introduction

Modern autonomous robots cannot be imagined without artificial intelligence systems, namely neural networks. The robot must solve strictly defined tasks, the complexity of the neural network must be very large and the learning time corresponding to the complexity of the process. The solution for this problem is multilayer convolutional neural networks. Such and similar tasks have been addressed by the authors in their scientific works [1, 2, 3, 4, 5]. The five main groups of methods for accelerating the training of convolutional networks exist

Methods

Results

Discussion

Conclusion