5812 South Ellis Ave
Chicago, IL 60637
Nathan Buerkle (Palmer Lab)
Title: "Ancestral nervous systems constrain the evolution of neural computation"
Abstract: Evolving new behaviors or perceptual abilities requires modifying the complex organization of neural circuits. This ancestral circuitry might impose significant constraints on which behaviors can feasibly evolve as well as the specific computations implemented. To address questions about the evolution of neural computation, we used machine learning to model the repeated evolution of red-sensitive, tetrachromatic color vision in butterflies from a UV/blue/green trichromatic ancestor. We first trained networks with trichromatic inputs to discriminate color stimuli. To simulate color vision evolution, we then retrained these networks with ‘mutated’, tetrachromatic inputs. Examining network performance showed that these evolved networks perform as well as, or even better than, tetrachromatic networks trained de novo from random starting weights. We also analyzed the computational structure of the hidden layer using hierarchical clustering and show that trichromatic starting points restrict how tetrachromatic networks implement novel computations. Overall, our network simulations showed that ancestral neural circuits do constrain the evolution of new computations and have the potential to promote or inhibit particular evolutionary trajectories.