Stimulus search and optimisation for visual neurons
Conventional experimental methods involve finding effective stimuli to describe the selectivity of sensory neurons. This process is not easy even in low-level sensory areas, and becomes highly difficult and subjective in higher visual areas, such as V4 and IT. Earlier stochastic methods, such as reverse correlation, aim to give a more objective and automatic way of characterising neurons at the lowest levels, but are also limited by the context-sensitivity due to non-linearities. I will describe two novel approaches that can be applied to testing higher-level neurons: 1) stimulus winnowing using large sets of natural images, 2) stimulus optimisation using gradient ascent. I will describe these and illustrate their application to neurons in visual areas STSa and V1 complex cells, respectively.