Training highly effective connectivities within neural networks with randomly initialized, fixed weights

TL;DR

This paper introduces simple, hyperparameter-free methods for training neural network connectivities with fixed, randomly initialized weights, achieving high performance with minimal connection modifications and shedding light on network over-parameterization.

Contribution

The authors propose novel methods for training neural connectivities without weight updates, eliminating hyperparameter tuning and reducing the effective size of networks.

Findings

Achieve over 90% of standard accuracy by flipping less than 10% of connections.

Match or surpass performance of trained weights with fixed, random weights.

Effective training of connectivities even with constant or asymmetric weight distributions.

Abstract

We present some novel, straightforward methods for training the connection graph of a randomly initialized neural network without training the weights. These methods do not use hyperparameters defining cutoff thresholds and therefore remove the need for iteratively searching optimal values of such hyperparameters. We can achieve similar or higher performances than in the case of training all weights, with a similar computational cost as for standard training techniques. Besides switching connections on and off, we introduce a novel way of training a network by flipping the signs of the weights. If we try to minimize the number of changed connections, by changing less than 10% of the total it is already possible to reach more than 90% of the accuracy achieved by standard training. We obtain good results even with weights of constant magnitude or even when weights are drawn from highly asymmetric distributions. These results shed light on the over-parameterization of neural networks and on how they may be reduced to their effective size.