|Ph.D Student||Shimron Avner|
|Subject||Efficient Estimation of Encoding Models from Visual Cortical|
|Department||Department of Biomedical Engineering||Supervisors||Professor Amit Meller|
|Professor Shy Shoham|
A central problem in neuroscience is to describe mathematically the encoding relations between incoming stimuli and the resulting neural activity. Encoding models range in complexity and accuracy depending on the type and size of the stimulus and the intended application; they help to extend our understanding of neural information processing and representation and to develop new sensory neuroprosthetics. Receptive field (RF)-type encoding models in the visual system, traditionally estimated using reverse-correlation methods, have provided much of our knowledge about the early stages of visual processing. However, recent developments in functional neural imaging challenge these techniques both by providing an unprecedented amount of data (often recorded simultaneously from hundreds of neurons) and by posing new experimental restrictions on the stimulus set. Moreover, estimation of simplified encoding models from a narrow set of visual inputs may miss basic properties and limit the ability to predict neural responses to novel stimuli.
Here, I use two different strategies to develop and demonstrate effective new methods for estimating mathematical encoding models of primary visual cortex neurons. In the first part, I describe how sparsity-based tools such as compressive sensing (CS) may be used to solve the inverse problem of RF mapping (RFM). I formulate an analogy between CS and RFM and demonstrate the effectiveness of CS-based mapping using both simulation and in-vivo visually evoked neuronal responses (made available by the Allen Institute). I demonstrate how CS-RFM provides information not attained by standard methods.
Next, I tackle the neural blind-identification problem in which one aims to infer information about a neural system without the ability to directly control or record its input. I provide two examples for this general approach: (a) using the first order statistics and (b) using second-order pairwise correlations. In example (a), I show how a static nonlinear function routinely used in neural encoding models may be estimated without direct knowledge of the stimuli, but with only its long-term statistics. The new technique uniquely estimates, using a maximum-likelihood approach, a physiologically plausible parametric model, given only the stimuli long-term statistics and observed response; that is, without requiring direct knowledge of the stimuli. In example (b), I show how spatial properties of RFs may be inferred from the population activity long-term pairwise correlations.
The new approaches could facilitate efficient estimation of neural encoding models from experiments where the input is highly controlled or from new realistic recordings where the stimuli are not easily monitored.