Friday, May 20, 2016

Autism, Artificial Neural Nets, and Language

It would benefit people to learn how artificial neural nets (ANNs) work in order to better understand the autistic brain. One of the things I've noticed in my readings on autism is that the way autistic people learn strongly resembles the way one trains up ANNs. It takes many iterations of a perception for the autistic person to develop the concept, whereas with neurotypicals only one or two will do.

Neurotypical people learn language in a more language instinct-driven fashion. A word needs be repeated only a few times, and the person has it. But if you tried to train an ANN, you would find that you would need to train it up many more times, and more than that, the first thing it would do is simply repeat back exactly what you said to it. That is, it would engage in echolalia.

Language is thus structured differently in the autistic brain than in the neurotypical brain. The neurotypical brain has a deep grammar on which the details of a given language are  hung. If autistic brains lack in certain kinds of instincts, like the language instinct, but still have enough complex network structure to build language, then we would expect less-than-typical structures in speaking. The unusual intonations of many autistic people likely derive from this fact as well, since every sentence is being constructed in a more mechanical way--giving us something like the voice of an ANN that learned language--rather than in the easier way generated by an instinct.

So language, in those who can learn it, is one of those many social behaviors neurotypical children learn automatically, but which requires direct instruction for autistic children.

No comments :

Post a Comment