Using the connectomic reconstruction of the optic lobe, we built neural networks that encode the sparse and hierarchical structure of the fly’s visual system, aiming to transfer inductive biases evolved in nature into artificial systems. Our approach feedforwardizes the recurrent structure of the optic lobe while preserving biologically meaningful processing sequences. These networks exhibit distinctive patterns of skip connections, favorable training dynamics, and competitive performance. We established a method for approximating convolutional filters from biological connectivity, enabling initialization schemes that encode evolutionary priors rather than relying on random weights.