Ancestral neural computations constrain the evolution of novel computations in simulated color vision networks

Efficient coding has been a successful organizational principle in neuroscience, but a more general theoretical framework needs to include the capacity for biological constraints to impede the realization of optimal circuit design. Here, we explore how evolution shapes the computational organization of a circuit using color vision as a model system. Taking a theoretical, machine learning approach allowed us to simulate the evolution of tetrachromatic color vision from a trichromatic ancestor both within and across distinct phylogenetic lineages. Analyzing network performance showed that trichromatic starting weights impose a significant constraint on learning rate, although the incremental increase in input layer complexity leads to better overall performance. Analyzing hidden layer computations showed that ancestry severely constrained evolution into a restricted and predictable portion of the theoretically available computational state space. Overall, our simulations of color vision evolution suggest that phylogenetic history is an important aspect of the functional organization of neural circuits.