Bounded Rational Decision-Making in Feedforward Neural Networks

TL;DR

This paper introduces a novel information-theoretic regularization approach for multilayer feedforward neural networks, inspired by bounded rational decision-making, which effectively prevents overfitting and achieves competitive results on MNIST.

Contribution

It applies bounded rationality formalism to neural network training, deriving new weight update rules that serve as regularization techniques.

Findings

Information-theoretic regularization prevents overfitting.

Achieves competitive accuracy on MNIST.

Applicable to various neural network architectures.

Abstract

Bounded rational decision-makers transform sensory input into motor output under limited computational resources. Mathematically, such decision-makers can be modeled as information-theoretic channels with limited transmission rate. Here, we apply this formalism for the first time to multilayer feedforward neural networks. We derive synaptic weight update rules for two scenarios, where either each neuron is considered as a bounded rational decision-maker or the network as a whole. In the update rules, bounded rationality translates into information-theoretically motivated types of regularization in weight space. In experiments on the MNIST benchmark classification task for handwritten digits, we show that such information-theoretic regularization successfully prevents overfitting across different architectures and attains results that are competitive with other recent techniques like dropout, dropconnect and Bayes by backprop, for both ordinary and convolutional neural networks.