Evolutionary Synthesis of Deep Neural Networks via Synaptic Cluster-driven Genetic Encoding

There has been significant recent interest towards achieving highly efficient deep neural network architectures that preserve strong modeling capabilities. A particular promising paradigm for achieving such deep neural networks is the concept of evolutionary deep intelligence, which attempts to mimic biological evolution processes to synthesize highly-efficient deep neural networks over successive generations that retain high modeling capabilities. An important aspect of evolutionary deep intelligence that is particular interesting and worth deeper investigation is the genetic encoding scheme used to mimic heredity, which can have a significant impact on the way architectural traits are passed down from generation to generation and thus impact the quality of descendant deep neural networks. Motivated by the neurobiological phenomenon of synaptic clustering, where the probability of synaptic co-activation increases for correlated synapses encoding similar information that are close together on the same dendrite, we introduce a new genetic encoding scheme where synaptic probability within a deep neural network is driven towards the formation of highly sparse synaptic clusters. Experimental results for the task of image classification demonstrated that the synthesized `evolved' offspring networks using this synaptic cluster-driven genetic encoding scheme can achieve state-of-the-art performance while having network architectures that are not only significantly more efficient (with a ~125-fold decrease in synapses at a comparable accuracy for MNIST) compared to the original ancestor network, but also highly tailored for GPU-accelerated machine learning applications.