Learning Compact Neural Networks Using Ordinary Differential Equations as Activation Functions

TL;DR

This paper introduces differential equation units (DEUs), a novel neuron activation approach allowing each neuron to learn its own nonlinear function during training, resulting in more compact networks with competitive performance.

Contribution

It proposes DEUs, enabling neurons to adapt their activation functions via ODE solutions, improving network efficiency and performance.

Findings

DEUs lead to smaller networks with comparable accuracy.

Neurons dynamically adapt their activation functions during training.

Improved network compactness without sacrificing performance.

Abstract

Most deep neural networks use simple, fixed activation functions, such as sigmoids or rectified linear units, regardless of domain or network structure. We introduce differential equation units (DEUs), an improvement to modern neural networks, which enables each neuron to learn a particular nonlinear activation function from a family of solutions to an ordinary differential equation. Specifically, each neuron may change its functional form during training based on the behavior of the other parts of the network. We show that using neurons with DEU activation functions results in a more compact network capable of achieving comparable, if not superior, performance when is compared to much larger networks.