This paper and the application program developed, will present my work in the field of Automated Learning. Starting from a philosophical discussion, I developed the idea of a neuronal device (a very vague model of a neuron; from now on referred to as MneMonic device). In my effort to find whether that model -- or a similar one -- already exists, I encountered results which further improved the practical basis of my model. By applying heuristics in all the three steps of the modelling (problem consideration, design and implementation) and by keeping a distance from the "too literal" models which seem to be preferred in the field, I implemented a neural net which can be trained to communicate in any language and then use the fuzzy knowledge base generated by this training to evaluate the correctness and veracity of a statement or the accuracy of a theory. The model uses genetic rules (mutations at threshold values and cross-overs in case of contextual vicinity) and a generator of "questions" (whenever a particular subnet is not logically balanced), both implemented at the highest abstraction level of the computatonal object, together with the classical synapse-creation, and "firing" methods.
The structure of the MneMonic device is very dynamic -- as the net "learns", it adds not only new instances of the same type of neuron, but also new types of neurons. In the process of creation of new neurons I involve a balance between the potential chaos induced by the genetic methods and plain logical operators. Actually one of the levels of neurons -- the meta- level, which takes charge of the internal development of the net -- increases considerably the heuristic operation of the resulting net. This guarantees a better solution to a problem in a smaller time interval.
In the current application, given the time constraints, I wire-strapped at the core of the net a grammar parser for the English language, so that visible results can be obtained faster. I used a syntactic model of the English grammar, developed by Prof. Bogdan Atanassov, but the semantic activity of the net remains to be derived through learning. I will be satisfied if the experiment will show that MneMonic can solve problems correctly when dealing with such a fuzzy input set that the human language is.
Last modified: 25Mar1998
Interestingly enough, I have not worked specifically on this project ever since I started studying Cognitive Science in earnest. I was always too busy reading about other people's models, working on other people's projects, working through the ups and downs of my family and generally staying alive. Pretty soon, this may change, though, as I elaborate in my new dissertation proposal.
Here are two frustrated bits of: poetry(?) and dizzy commentary about MneMonic and rotten chunks of life in general.
|[ Home ][ Current Projects ][ Portfolio ][ Pastime ][ Pseudonyms ][ CV ]|