Abstract:
As animals interact with their environments, they need to cope with considerable ambiguity and uncertainty inherent in their sensory input. Recent psychophysical experiments suggest that at some level the brain does this by implementing Bayesian inference. This thesis considers algorithms from Machine Learning which perform Bayesian inference on graphical models, and investigates how these algorithms may be implemented with neural components. In this view, neurons are the variable nodes of a graphical model and synapses represent the dependencies between variables. These algorithms are based on the propagation of messages that convey probabilities. Furthermore, learning of the probabilistic dependencies between variables of a belief network is based on observed data and performed in an online and unsupervised manner using a variation of the expectation-maximization (EM) algorithm. The feasibility of these algorithms is demonstrated on a motion detection task.
