The International Computer Science Institute
is pleased to present a talk:

"Neural Network Training for Phoneme Recognition using Evolutionary Algorithms: Preliminary Results"

Konsta Koppinen ICSI
konsta icsi.berkeley.edu

Tuesday, August 10, 2004 ICSI, Conference Room 5A 12:30 pm

This talk will discuss the training of neural networks for phoneme classification using evolutionary algorithms instead of backpropagation. That is, the normal MFCC features of a frame of speech go into the network and the network has graded outputs for each of the phonemes. The network is grown randomly by trying out different mutations (adding/deleting neurons, adding/deleting inputs to the neurons, changing the weight/bias of neurons) and evaluating the results. Thus the topology of the network is determined along with the weights and the resulting network will (most likely) be sparse.

The current state of the system I have implemented will be described, including: a demonstration of how the evolutionary algorithm builds a correct feedforward neural network for a simple task. For efficiency, the system estimates the value of a given neural network from its operation on just a few samples and a simple mechanism for weeding out the unpromising networks will be described. A few examples of how different types of networks can be built by different choices of evaluation function are given. Preliminary phoneme recognition results with the TIMIT database will also be shown.