Abstract
What can we learn from a connectome? We constructed a simplified model of thefirst two stages of the fly visual system, the lamina and medulla. Theresulting hexagonal lattice convolutional network was trained usingbackpropagation through time to perform object tracking in natural scenevideos. Networks initialized with weights from connectome reconstructionsautomatically discovered well-known orientation and direction selectivityproperties in T4 neurons and their inputs, while networks initialized at randomdid not. Our work is the first demonstration, that knowledge of the connectomecan enable in silico predictions of the functional properties of individualneurons in a circuit, leading to an understanding of circuit function fromstructure alone.