Benjamin Cowley

The brain is as fascinating as it is complex: When an image lands on the retina, our brain picks out relevant objects and salient details by using its massive network of interconnected neurons. What are the computations lurking within this network?

To understand the brain’s computations, one often seeks a model to capture the step-by-step computations of neurons. For example, a model can take an image as input and output a visual neuron’s response. The most predictive models in computational neuroscience typically have millions of parameters and need large amounts of training data—making it difficult to obtain and interpret such models.

The Cowley research group takes a two-pronged approach to address these problems. First, we design adaptive stimulus selection techniques to efficiently train models (e.g., deep neural networks) with as little recording time as possible. We work hand-in-hand with experimentalists to deploy these systems in closed-loop experiments. Second, we develop machine learning techniques to identify highly-predictive models that have as few parameters as possible. We then analyze these “compact” models to determine the computations necessary and sufficient to explain a neuron’s response. Our approach of identifying highly-predictive, interpretable models will shed light on computations otherwise hidden by the scale and complexities of the brain.