Evolving Simple Solutions to the CIFAR-10 Benchmark using Tangled Program Graphs

The goal of the CIFAR-10 benchmark is recast from the perspective of discovering light-weight as well as accurate solutions. Specifically, the image data, on which CIFAR-10 is based, requires multiple practical issues to be addressed that are not often considered collectively when applying genetic programming to classification problems. Issues of particular interest include cardinality, multi-class classification and diversity maintenance. We demonstrate that diversity maintenance and cardinality can be approached simultaneously by adopting a data subset to compose pools of exemplars for lexicase selection. The issues of multi-class classification and solution simplicity are addressed by adopting the tangled program graph (TPG) approach to emergent modularity. In addition, the mutation operator is modified to ensure that class labels do not ‘die out’ during evolution. The resulting benchmarking study demonstrates solutions that are significantly more accurate than AutoML while providing comparable accuracies with solutions from unsupervised feature discovery, i.e. 70% accuracy. However, unlike the latter TPG solutions are several orders of magnitude simpler.