Friday, 2 November 2012

// // Leave a Comment

Noam Chomsky: Where AI went wrong

Noam Chomsky was interviewed by the Atlantic. The article is well written but biased towards Chomsky's viewpoint. I think the title of the article is ridiculous, but that's the modern media for you. It's inflammatory, over-reaching and ignorant. AI may have fractured but it's not gone wrong. Gone wrong is when your toaster decides your dog needs to die. I do like Chomsky's viewpoint that:
"new AI" -- focused on using statistical learning techniques to better mine and predict data -- is unlikely to yield general principles about the nature of intelligent beings or about cognition
Which is quite true and has been eating away at me for some time. My dynamic neural networks, although used for clustering and classification, have the potential to simulate more organic systems. The way I actually treat my algorithms is without due care. I'm less bothered with finding maxima but investigating the dataset to see what there is. Novelty and discovery are more important to me than function approximation or reasoned logic. My latest batch of code (written in javascript) has the facility to give the network neuron damage to see what the affect on learning is. Optimisation it certainly isn't. It's a child-like approach to investigation (and mathematics!). What the article did for me was hold up a mirror. Some years have passed since graduating with my PhD, I'm still fascinated by my wobbling algorithms but have struggled to grasp why. Now I think it might be that I am more interested in the biological parallels than using them to predict, cluster or classify. The parallels are bi-directional. Steal more from biology, understand a little more.

Action

With that in mind, I can see my own research with greater lucidity. It's OK not to worry about something's ability to classify or cluster, what other biological parallels can I draw. Firstly, I think neurons with different properties (rather than homogeneous ones) might work well, as might limits on connectivity. That would better represent the physical limits of the real world. I'm very much thinking out loud now, which leads me into trouble.