Final Project progress

My final project is exploring an unsupervised neural net in the framework of stereotyping. Unsupervised neural nets just search for patterns on the data that has been entered and does a sorting just based on similarities of certain traits. It doesn't require a teaching signal and it seems to do a decent job of sorting based on a simple REALbasic construction of it that I have used. I haven't yet been able to get my implementation to compile when adding the stereotyping component to it, but I'm still hopeful. I'm using a network based on Stephen Grossman's work: the Adaptive Resonance Theory network. It uses the vigilance parameter to yield formation of new nodes. I expand it by allowing these nodes to backpropagate weights through the entire network. This backpropagation allows there to be change of the original stereotypes.

I've been reading a lot on a couple of different network models used in social cognition. First, the tensor product model which uses the tensor product of two matrices and has the interesting property of being able to add context to an event. I may try to explore using the basics of the tensor product model in implementing my new network. My network uses vectors (in the linear algebra sense) as the basis for my input patterns. So, I can just take the tensor product of the vectors used as my input data and look at the kind of relationships it can build.

What's interesting about the model I'm building is the way it clusters the pattern representations. Essentially, it sorts the vectors by seeing if they're "closest" to each other in n-dimensional space. It's easy to visualize if the vector has 3 or less members of its pattern set, but beyond that we enter fourth dimension and into difficult to imagine scenarios.

Anyway, this model has been very interesting to explore and I hope it can yield some usable results.