Using Statistics of Natural Images to Facilitate Automatic Receptive Field Analysis

A neuron can be conceptualized as performing a function on its inputs. This stimulusresponse function characterizes “what the neuron does” and is the main object of interest for cellular-level neuroscience. Automated methods for discovering certain properties of stimulus-response functions have many advantages over traditional methods [5]. They can perform more exhaustive searches, use less-biased inputs and have tighter feedback loops for online, adaptive analysis. However, automated methods require many stimulus presentations, often many more than are feasible for a given physiology experiment. Finding ways to reduce the number of stimulus presentations is crucial for increasing the power and scope of these important techniques. We believe that one way to do this is to take advantage of the statistics of natural images. Neurons in V1 appear to be tuned to the specific statistical properties of natural images. For example, a variety of techniques like sparse coding [13, 7], independent component analysis (ICA) [1, 8] and slow feature analysis (SFA) [17, 2], when applied to natural images, yield response properties strikingly similar to both simple and complex cells in V1. That neurons are tuned to natural inputs makes sense from both an evolutionary perspective and from a developmental perspective. We can use this knowledge to more efficiently probe the response properties of visual neurons. A simple thought experiment partially illustrates our idea. Natural images can be significantly compressed using algorithms like JPEG without suffering any noticeable loss in quality. This suggests that neurons in the visual cortex ignore certain types of variation in images. If we want to somehow search for the stimulus that makes a neuron