Tuesday, July 29, 2008

Neural Computation and Psychology Workshop 11

The eleventh Neural Computation and Psychology Workshop (NCPW11) was held this year at the Experimental Psychology department of Oxford University in historical Oxford on July 16-18. There have been several exciting presentations in the workshop, of which I summarizing a few I found most interesting.

Dr. L. Andrew Coward (Australian National University), whom I had earlier met at AKRR'05 gave an interesting talk about managing neocortex resources. In his talk, Dr. Coward told us about a his artificial neural network model that uses leaky integrators. The structure of the ANN is based on the findings that the receptive fields of the cortical pyramidal cell columns in the brain seem to overlap as little as possible. According to him, the responses in the cortical columns form the information record which serves as a declarative memory, and the association of the columns in which there was an increase in the receptive field simultaneously is the the episodic memory. The background idea of minimally overlapping (i.e. as independent as possible?) features sound very much like Independent Component Analysis (ICA) to me, except that in ICA you cannot add more features to better discriminate between categories. Further talk with Dr. Coward clarified that the outcome of the model is indeed similar to ICA, but the matemathical foundations of the two approaches are fairly different. The simulations are carried out using random patterns in different categories making sure that the inputs of the same category are statistically similar.

Dr. John Lipinski (Institut für Neuroinformatik, Ruhr-Universität Bochum) gave another interesting talk. Test subjects show a so-called spatial drift in behavioral experiments in which they are shown a cross on the computer screen, and then after a short delay, their task is to click the place where the cross was. In the experiments, there is a considerable drift toward more left/right/above prototype depending on the place of the target. According to further experiments, the effect is even stronger when a lexical category label (above, left, right) is shown. Building on these findings, he then introduced a model for simulating the integration of linguistic and non-linguistic spatial systems using dynamic field theory approach to spatial working memory integrated with a competitive spatial semantic network. According to the results shown, the simulation model captures the empirical findings.

Dr. Nicolas Ruh (Oxford Brookes University) presented OXlearn, a new neural networks simulation toolkit. The toolkit contains simple neural network models, a graphical interface in which several parameters can be altered. The toolkit is intended to be an easy-access pedagogical tool, which means no programming is needed to use it (but you can access all the Matlab-based parameters and functions, if you want to). The OXlearn is available at http://psych.brookes.ac.uk/oxlearn for free download both as a toolbox for Matlab, and as a standalone version (only for Windows XP at the moment). Within the field of computer science, one usually wants to teach the students programming as well, but there are several other disciplines in which such a program is probably quite useful.

The invited speaker in this year's workshop was Professor David Plaut from Carnegie-Mellon University (Psychology, Computer Science and the Center of the Neural Basis for Cognition). He gave an insightful presentation about the common cognitive and neural principles of face and word processing in the brain. The research had been carried out together with professor Marlene Behrmann. Distributed presentation of knowledge as a pattern of activation over multiple, hierarchically organized visual areas. According to him, the representations are different in the case of faces and words, but the computation would be same and in both cases the process relies on high visual acuity which enables making fine grained discriminations. In addition, representational coordination and competition and a bias for local connections are important. His conclusions were based by simulations carried out in two simulated systems which on these features, and the results of the simulations seem to produce behavior which is similar to humans in these cases.

The last talk of the three-day workshop was given by Dr. András Lorincz (Eötvös Loránd University, Budapest, Hungary). This talk featured a cognitive model based on factored reinforcement learning and independent component analysis. Like most of the content of the workshop, I will look forward to reading the full paper of this paper the proceedings of the workshop that should be out in December.

The NCPW12 will be probably be held in London in early 2010. I am certainly looking forward to it!

No comments: