One of the hallmarks of the Allen Brain Observatory data is the trial-to-trial variability in neural responses. We analyzed the distribution of responses for each cell’s “preferred” stimulus within a repeated natural movie clip and found that many cells show non-Gaussian noise distributions. We repeated the experiment using electrophysiological recordings and found similar results. We also used a subspace analysis to show that the noise aligns with variations within a movie clip, but not across different movie clips. One possible role for this noise is to help cortex generalize, i.e. to help cortex learn general representations from a small number of examples. This has important implications for understanding computations in cortex, and may also have applications for training artificial neural networks with better generalization abilities.
Please see our new preprint!