Lamarckian Evolution of Convolutional Neural Networks

TL;DR

This paper demonstrates that Lamarckian evolutionary algorithms can efficiently optimize convolutional neural network architectures by inheriting learned weights, leading to faster convergence and improved data efficiency on image classification tasks.

Contribution

It introduces a Lamarckian evolution approach for CNN architecture search that inherits learned weights, reducing training time and improving performance.

Findings

Similar or better test accuracies compared to baselines.

Significant reduction in training data requirements, up to 75%.

Faster convergence speeds observed in experiments.

Abstract

Convolutional neural networks belong to the most successul image classifiers, but the adaptation of their network architecture to a particular problem is computationally expensive. We show that an evolutionary algorithm saves training time during the network architecture optimization, if learned network weights are inherited over generations by Lamarckian evolution. Experiments on typical image datasets show similar or significantly better test accuracies and improved convergence speeds compared to two different baselines without weight inheritance. On CIFAR-10 and CIFAR-100 a 75 % improvement in data efficiency is observed.