Investigating the interaction between gradient-only line searches and different activation functions

TL;DR

This paper explores how gradient-only line searches (GOLS) interact with various activation functions in neural networks, revealing robustness issues with ReLU and solutions via architectural modifications like batch normalization and skip connections.

Contribution

It investigates the impact of activation function properties on GOLS performance and demonstrates how architectural features can mitigate associated training difficulties.

Findings

GOLS are generally robust across activation functions.

ReLU's zero-gradient region causes sparsity issues affecting training.

Batch normalization and skip connections improve GOLS effectiveness with ReLU.

Abstract

Gradient-only line searches (GOLS) adaptively determine step sizes along search directions for discontinuous loss functions resulting from dynamic mini-batch sub-sampling in neural network training. Step sizes in GOLS are determined by localizing Stochastic Non-Negative Associated Gradient Projection Points (SNN-GPPs) along descent directions. These are identified by a sign change in the directional derivative from negative to positive along a descent direction. Activation functions are a significant component of neural network architectures as they introduce non-linearities essential for complex function approximations. The smoothness and continuity characteristics of the activation functions directly affect the gradient characteristics of the loss function to be optimized. Therefore, it is of interest to investigate the relationship between activation functions and different neural network architectures in the context of GOLS. We find that GOLS are robust for a range of activation functions, but sensitive to the Rectified Linear Unit (ReLU) activation function in standard feedforward architectures. The zero-derivative in ReLU's negative input domain can lead to the gradient-vector becoming sparse, which severely affects training. We show that implementing architectural features such as batch normalization and skip connections can alleviate these difficulties and benefit training with GOLS for all activation functions considered.