Analyzing neural responses to more natural stimuli

William Bialek

Most experiments on neurons are done with simplifed stimuli that are chosen from low dimensional parameter spaces. This approach has the attraction, that one can plot some measure of neural response against the few parameters, and of course much has been learned this way. If we leave low dimensional stimuli behind in favor of natural signals, what can we measure? One answer is that we can measure information transmission. I will summarize methods that allow us to measure mutual information between spike trains and stimuli for aribtrarily complex stimuli, provided only that we can repeat the stimuli many times. We can also measure the information carried by single spikes, bursts, coincident firing of cell pairs, ... , all in a model independent manner. These results provide a benchmark against which our understanding of the neural code can be tested. We want to know not just how much the spike train is telling us, but also WHAT it is telling us. Our intuition is that single cells are selective for features, and I will discuss a method for identifying candidate 'feature dimensions;' if there are just a few of these interesting dimensions then we can map out a fully nonlinear input/output relation in this low dimensional space. This is work in progress, so I don't quite know how much I'll be able to say, but I will certainly show how these methods have been used to reveal some striking adaptation to the statistics of velocity signals in the motion sensitive cells of the fly.
Back to the Natural Scenes Meeting Agenda.