Possible project themes for Neurocomputing 2006

Each theme should be researched by two students. While you are free to study the theme together and exchange ideas, every student is expected to produce a unique paper.

Biological self organized maps (SOM):

Self organizing maps have been popular in computer science as unsupervised clustering algorithm, specifically in the implementation proposed by Kohonen. However, it also forms the a basic model of cortical organization that I believe is central in many types of learning. The biological implementation of Willshaw and Von Der Malsburg is well worth to revisit.

Hetero-associative sequence memory (HNN)

There is increasing evidence that temporal aspects of memory are essential in human memory processing. Sequence memory was also discussed in many early papers of recurrent associative memories. We have recently proposed a modular recurrent model for sequence processing that bares some similarity to a synfire chains model, and further work ion these type of models look promising.

Storage capacity of sparse and noisy auto-associative networks (ANN)

We have illustrated a specific recurrent memory for information storage that used a maximal distributed representation of pattern. It is interesting to see how these networks work in more biological realistic settings such as sparse representation which we think are common in the hippocampus and also noisy processing.

Continuous Attractor Neural Network (CANN) for decision making

Decision making is a recent very active research area. Continuous attractor neural networks are a specific form of recurrent networks that have been used to model such processes. However, these models have been used in a particular mode and the thread of this project would be to investigate decision making in a new novel domain of this network. There is a recent special issue of Neural Network on decision making. Yhe main thrust of this project is that I have an idea that you should investigate, including comparison to literature.

Applications of Support Vector Machines (SVM) to specific application area

Support Vector Machines are the latest incarnation of perceptron-type machine learning methods. In this project you need to read up a bit on the theory and apply it to a specific application area (classification) of your choice. Some good tutorials are the ones by Burges and also Smola and Schoelkopf. Have also a look at the introductory resource page with links to imlementations by Matt Boardman.