I have worked on a few projects over the years. The first was a game playing program designed to make it easy to experiment with new game designs. The project report is here and code is available here and on sourceforge. The second project, described below was a computer program to play music with style and expression rather than the mechanical sound normally associated with electronic music. Next I shifted from music to linguistics, but continued to apply much the same techniques, and some of my results are described here. Most recently I have been trying to apply of the techniques I developed in linguistics into other fields such as product suggestions. There are no publications relating to this work, but I discuss it briefly below.
There are many other small projects I have experimented with but never turned into full publications. For instance, I developed a codec for extremely low bandwidth but high quality speech (similar to skype) based on sending a speech model at the start of the conversation. I have also been designing a Go AI on and off, but have not come up with anything that is demonstrably better than existing technologies.
My second project was in computer performance of piano music. Two reports are available, an interim and a final. The most interesting part of this project was the finding that good performance is entirely dependent on good data representations for the neural network. This project represented input in such a way as psychologically natural chords are natural and easy to represent while unnatural or impossible actions are hard or impossible to represent. Further details are included in the report, but this concept, that success or failure will be determined by the quality of the neural network's data format has been very useful in every other project I have worked on.
The similarity between music and language has been observed by many people, and it was quite natural to shift from grammars of music to grammars of English. I concentrated on developing a parser that was trained as much as possible on actual English usage rather than having any linguistic knowledge built in.
This research came up with a number of interesting findings. Enumerating them all here would be infeasible but the ones I have publications on are: a part of speech tagger that returns a probability distribution of possible tags rather than just the most likely tag; a statistical parser incorporating a number of highly efficient algorithms and data structures so as to achieve reasonable speed while still being easy to modify (details); a completely unsupervised model of language which can be used to test various linguistic theories (details); a probability model based on a neural network rather than counting numerators and denominators (details), making it possible to predict the probability of events when the model is too large for accurate counts. The thesis of this work is available here.
In my language research I develop a number of interesting techniques for automatically extracting knowledge, and also a neural network to simulate a statistical probability model. For my next research I would be interested in applying these techniques to other fields. For instance the knowlege extraction technique can be applied just as easily to a corpus of genetic information, and the neural probability model can be used anywhere that it is non-obvious what parameters the probability model should take.
Some of the particular applications that have interested me are product recommendation systems, such as that used by amazon, and document classification -- perhaps using tools developed in spam filtering to help with technical support.