טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentKatz Shahar
SubjectA Bayesian View of Spontaneous Neural Activity as an
Internal Representation of the Environment - a
Neural Network Model
DepartmentDepartment of Electrical Engineering
Supervisor Professor Ron Meir
Full Thesis textFull thesis text - English Version


Abstract

_____________________________________________________________________________


Studies in the past decade have demonstrated increasing success in identifying and reconstructing images and video clips from neural activity. The success of these schemes relies heavily on applying prior knowledge about the inherent attributes of natural scene stimuli. The combination of prior knowledge and activity-dependent statistical modeling suggests that the perception process that occurs in the visual cortex is essentially Bayesian. A Bayesian model of perception constructs priors over plausible stimuli and computes posterior probabilities given sensory input.


Findings in recent years have concluded that spontaneous or ongoing neural activity plays a major role in perception and development. Numerous studies have attempted to capture characteristics of spontaneous activity and its mode of integration with sensory input. These characteristics include resemblance to evoked activity, slow oscillations between “up” and “down” states, low variability etc. It is known that recurrent activity in the Visual Cortex, occurring even in the absence of external stimuli, dominates thalamic input from the senses in the visual cortex by an order of magnitude. Therefore, spontaneous activity is a plausible candidate to implement the Bayesian priors that reflect the inherent knowledge about the structure of natural stimuli that is gathered through visual experience. In this scheme, the input stimulus is integrated with spontaneous activity to compute posterior probabilities over the object being perceived.


Recent studies have developed computational models aiming to incorporate spontaneous activity in artificial neural network models. Our study suggests a network model consisting of Fitzhugh-Nagumo neurons which utilizes spontaneous activity to construct priors and implement Bayesian inference given input serving as a bifurcation in the state space dynamics. The proposed model displays spontaneous activity with similar features to those observed the literature, and is consistent with a Bayesian view of perception.