Training without Gradients -- A Filtering Approach

TL;DR

This paper proposes a novel gradient-free training method for neural networks using particle filtering, modeling weights as state variables and neural networks as measurement functions, demonstrated on a simple example.

Contribution

It introduces a filtering-based approach to train neural networks without gradients, offering an alternative to traditional backpropagation methods.

Findings

Preliminary demonstration on a simple example.

Potential for training deep networks remains to be explored.

Framework models weights as state variables in a filtering context.

Abstract

A particle filtering approach is suggested for the training of multi-layer neural networks without utilizing gradients calculation. The network weights are considered to be the components of the estimated state-vector of a noise driven linear system, whereas the neural network serves as the measurement function in the estimation problem. A simple example is used to provide a preliminary demonstration of the concept, which remains to be further studied for training deep neural networks.