Google Analytics

Wednesday, October 2

Aritificial Alien Intelligence

Like fashion, even scientific research is subjected to cyclical fads. Topics, techniques come into limelight and then fall out of favor when the gains plateau. 

In machine learning, NNs were in vogue in early 90s when SVM came onto the scene and blew them out of water on almost every problem and then got superseded in turn by the log-linear models, CRFs and others. Then in beginning of this decade, NNs have made a tremendous come back, this time becoming the tool of choice for NLP as well.

Due to their name and a very crude similarity to how the biological brain is organized, NNs tend to invite comparisons with human brain and every improvement achieved with their help tends to be seen as an step towards Artificial General Intelligence, at least in popular media and common people. 

Now, what would happen if tomorrow a different method, a different paradigm of Machine Learning were start performing better than NNs? There is no reason why it could not happen AFAIK. We do not have theoretical results that say that NNs are better then every other method imaginable. They are as capable as anything else possible but that only means that another method that is better at exploiting the given data may be able to beat them.

Would such a development still feel like a step towards AGI? When you hear about a CRF, it merely sounds like a fancy mathematical object which we have devised. We don't feel threatened by it. It is because NNs are presented in the form of neurons and similarity to biological brains that the question seems to have assumed this sense of urgency. Or is it because for the first time, these systems have started to come close to human performance in certain tasks?

How do you think the discourse will change if a method that cannot be presented as interconnected neurons firing takes the lead? Would we call it Artificial Alien Intelligence? Or we are now past the possibility of such reversals or upheavals and the next jump up will come within the same paradigm only?

No comments: