SAIL Room, 111 Levin Building (425 S. University Ave.)
Kording Lab, UPenn
Machine learning for problems of neural coding
A common goal within neuroscience is to identify which factors are informative of the activity of single neurons – that is, what information those neurons encode. This two-part talk will focus on the value of machine learning methods for this endeavor. Using test data from three brain regions, I'll first show that machine learning methods are often much more accurate than generalized linear models (GLMs) at predicting spike rates. Machine learning methods can act as performance benchmarks to check against the possibility of a GLM mischaracterizing nonlinearity, or, when one needs only predictive ability, as replacement methods. In the second part of the talk, I will discuss our ongoing effort to use natural image responses of macaque V4 neurons to extract color tuning curves. By comparing these curves to those built more traditionally from responses to images of a single hue, we wish to investigate whether color modulation is static or context-dependent. This requires controlling for the numerous other visual features that modulate V4 firing rates but that are difficult to parameterize (such as texture and curvature), making this a good application for machine learning methods. Our approach is to query the first-order color response of a successful encoding model built from a pretrained convolutional neural network. I’ll conclude with a brief overview of our open-source Python repository for machine learning encoding models.
Lunch will be served.