Improving Generalisation in Deep Learning through Quality Diversity

Deep learning has shown impressive results lately, not only because of new algorithmic inventions but also an substantial increase in computational resources. In this proposal we aim to scale up another method to train neural networks called neuroevolution. Neuroevolution has so far only been applied to problems and networks that are much smaller to current state-of-the-art deep learning methods. The hypothesis in this proposal ist that similarly to deep learning, the true potential of neuroevolution could be unlocked by scaling to significantly larger networks with 10+ million parameters.