By Ira L. Cohen, Ph.D.

 

I'm a psychologist studying children with autism and trying to better understand, treat, and possibly help prevent this condition. Studying neuroscience in graduate school, my post-doctoral work was in clinical experimental psychology. While I don't consider myself a neural networks expert, I probably know more about the subject that most in my field - thanks in part to some excellent NeuralWare courses I took quite a while ago.

When first using NeuralWorks Professional II/PLUS, back in 1994, I published (in the Journal of Autism and Developmental Disorders) an article about using neural networks as a way of classifying children with autism and discriminating them from other groups. In the process of so doing, we demonstrated that neural networks technology was superior to discriminate analysis in classifying the children. At that time, the field of diagnosing autism was diffuse; since then, so-called "gold standard" methods of diagnosing kids have been developed. While we can now reliably diagnose autism using these tools, our ability to define sub-groups is still problematic and we therefore continue to have interest in using neural network classifiers to help in this area.

Some years later, I realized that, neural networks and autistic children have something in common. They are both brilliant at some tasks, and abysmal at others. In fact, in the mid 1990's, the first good autopsy studies were coming out, showing that the brains of autistic people develop differently - they had too few neurons in some areas, and a much higher-than-average concentration in others. I began to see analogies between the way neural networks learn and the way autistic children learn. If you give either too little or too much information, they perform poorly. I used my copy of Professional II/PLUS to simulate autistic behavior, intentionally sub-optimizing the input and the design of the network. For example, I was varying the size of the hidden layer, and giving the application more and more complex problems in order to study rates of acquisition and rates of generalization. The correlation between the difficulty that poorly defined neural networks have with generalization and the difficulty that children with autism have with generalization proved to be eerily similar, and I published an article about the correlation. This helped me to understand what's wrong with the kids and why they learn the way they do. It also suggested ways of intervening.