Sequential training algorithm for neural networks

TL;DR

This paper introduces a sequential training algorithm for large-scale neural networks that trains each layer separately to reduce computational resources and improve convergence, offering a practical alternative to full network training.

Contribution

The paper proposes a novel sequential training method for neural networks that simplifies training and reduces resource requirements with minimal modifications to existing algorithms.

Findings

Reduces training time significantly

Achieves comparable performance with full training

Requires minimal changes to existing algorithms

Abstract

A sequential training method for large-scale feedforward neural networks is presented. Each layer of the neural network is decoupled and trained separately. After the training is completed for each layer, they are combined together. The performance of the network would be sub-optimal compared to the full network training if the optimal solution would be achieved. However, achieving the optimal solution for the full network would be infeasible or require long computing time. The proposed sequential approach reduces the required computer resources significantly and would have better convergences as a single layer is optimised for each optimisation step. The required modifications of existing algorithms to implement the sequential training are minimal. The performance is verified by a simple example.