Evolution of Activation Functions: An Empirical Investigation

TL;DR

This paper introduces an evolutionary algorithm to automatically discover new activation functions for neural networks, demonstrating improved performance over existing functions across multiple datasets and architectures.

Contribution

It presents a novel evolutionary approach to automate the search for activation functions, a task traditionally done manually or with predefined options.

Findings

Evolved activation functions outperform standard functions in experiments.

The approach is statistically robust across 10 datasets and 30 runs.

New activation functions show promising results for neural network performance.

Abstract

The hyper-parameters of a neural network are traditionally designed through a time consuming process of trial and error that requires substantial expert knowledge. Neural Architecture Search (NAS) algorithms aim to take the human out of the loop by automatically finding a good set of hyper-parameters for the problem at hand. These algorithms have mostly focused on hyper-parameters such as the architectural configurations of the hidden layers and the connectivity of the hidden neurons, but there has been relatively little work on automating the search for completely new activation functions, which are one of the most crucial hyper-parameters to choose. There are some widely used activation functions nowadays which are simple and work well, but nonetheless, there has been some interest in finding better activation functions. The work in the literature has mostly focused on designing new activation functions by hand, or choosing from a set of predefined functions while this work presents an evolutionary algorithm to automate the search for completely new activation functions. We compare these new evolved activation functions to other existing and commonly used activation functions. The results are favorable and are obtained from averaging the performance of the activation functions found over 30 runs, with experiments being conducted on 10 different datasets and architectures to ensure the statistical robustness of the study.