Arend Hintze, P. Robin Hiesinger, Jory Schossau

pdf

Abstract: Artificial neural networks (ANNs) learn through iterative ad- justment of synaptic weights, most commonly by backprop- agation or similar training methods. By contrast, learning in the neuro-evolution approach of ANNs as well as in biolog- ical neural networks can also be achieved through iterative adjustment of synaptic weights based on mutational changes of a genome. Evolutionary learning of ANNs is typically based on changes in the genome that directly affect synap- tic weights, while the biological genome affects changes to synaptic weights through indirect and developmental encod- ing of the molecular composition of synapses. Methods for indirect and developmental encoding in ANNs are compara- bly scarce and their constraints and advantages remain in- completely explored. Here, we identify robustness to en- vironmental challenges as a key feature associated with in- direct and developmental encoding using a neuro-evolution approach. In biological systems, the developmental process ’unfolds’ synaptic connectivity information in a time- and energy-consuming process that is sensitive to environmental conditions like temperature. Our tests for robustness to devel- opmental temperature reveal successful adaptation for indi- rect encoding followed by a developmental process compared to lower performance of developmental or indirect encoding alone. These findings set the stage for a more comprehensive neuro-evo-devo approach that acknowledges the necessity of neural network development for biological intelligence.

Part of the 2020 Workshop on Developmental Neural Networks