Joint spatial and temporal structure learning for task based control

We present an example of a joint spatial and temporal task-learning algorithm that results in a generative model that has applications for on-line visual control. We review work on learning transformed mixture of Gaussians (due to Frey and Jojic) and variable length Markov models (VLMMS due to Ron, Singer and Tishby). We show how a temporal model, learned through an extension of VLMMs to deal with multinomially distributed input symbol vectors, can be used as an improvement on maximum likelihood (ML) for prior parameter estimation for the expectation maximisation (EM) process.