Phase transitions in the generalization behaviour of multilayer perceptrons: II. The influence of noise

TL;DR

This paper investigates how different types of noise affect phase transitions in the generalization behavior of multilayer perceptrons, revealing that many phase transition features persist and can be modulated by noise levels.

Contribution

It extends previous work on phase transitions in perceptrons by analyzing the impact of input and output noise, including effects on optimal noise levels for learning.

Findings

Phase transitions mostly persist despite noise.

Noise effects can be modeled by rescaling the loading parameter.

Optimal noise levels depend on the learning paradigm.

Abstract

We extend our study of phase transitions in the generalization behaviour of multilayer perceptrons with non-overlapping receptive fields to the problem of the influence of noise, concerning e.g. the input units and/or the couplings between the input units and the hidden units of the second layer (='input noise'), or the final output unit (='output noise'). Without output noise, the output itself is given by a general, permutation-invariant Boolean function of the outputs of the hidden units. As a result we find that the phase transitions, which we found in the deterministic case, mostly persist in the presence of noise. The influence of the noise on the position of the phase transition, as well as on the behaviour in other regimes of the loading parameter $\alpha$, can often be described by a simple rescaling of $\alpha$ depending on strength and type of the noise. We then consider the problem of the optimal noise level for Gibbsian and Bayesian learning, looking on replica symmetry breaking as well. Finally we consider the question why learning with errors is useful at all.