hide
Free keywords:
-
Abstract:
Perceptual inference relies on very nonlinear processing of high-dimensional sensory inputs. This poses a challenge as the space of possible nonlinearities is huge and each of there functions might be implemented in many different ways. To gain better insights into the nonlinear processing of sensory signals in the brain, we are trying to make progress on two fronts: (1) By learning probabilistic representations of natural images, we explore which nonlinearities are most successful in capturing the degrees of freedom of the visual input. (2) We develop neural system identification methods with the aim of identifying unknown nonlinear properties of neurons that are difficult to identify by linear or generalized linear models. In this talk, I will first give a short summary of our results on natural image representations and our conclusions regarding effective nonlinearities (15min). In the second part (30min), I will talk about a system identification approach based on a new model called the spike-triggered mixture (STM) model. The model is able to capture complex dependencies on high-dimensional stimuli with far fewer parameters than other approaches such as histogram-based methods. The added flexibility comes at the cost of a non-concave log-likelihood but we show that in practice this does not have to be an issue. By fitting the STM model to spike responses of vibrissal afferents we demonstrate that the STM model outperforms generalized linear and quadratic models by a large margin (up to 200 bits/s).