The Believers: The story behind deep learning | Hacker News Reader

The Believers: The story behind deep learning

chronicle.com

42 pointsmjohn6 comments

Hinton gave a well attended keynote at the recently concluded AAAI 2015 conference in Austin, in which one of his concluding slides simply stated in a headline font ...

"GOFAI is finished"

I'll explain the irony of this dramatic moment. The AAAI is the Association of the Advancement of Artificial Intelligence. Its annual conference is the worldwide gathering of university graduate students, and a few corporate labs, presenting their research in the disparate niche areas that have something to do with smart or human like behavior. All these niche areas have their own annual conferences,... Robotics, Planning, Semantic Web, Knowledge Representation, Natural Language Processing, pattern recognition, speech processing, theorem proving, and so forth.

GOFAI is "Good Old Fashioned Artificial Intelligence", and is used to represent the line of mostly symbolic AI research made popular by John McCarthy, Marvin Minsky, and others back in the 1960-1970 era. It was those folks who as peer reviewers made Hinton's life awful. The NIPS conferences over the years became the home for connectionist researchers, as opposed to the AAAI.

Now the wheel has turned full about. The triumph of Hinton's approach allows him to tell thousands of young researchers that their own research approaches are finished unless they incorporate deep learning.

It's sad to think that the NN researchers will overreach and suppress other research approaches now that they are having a (well deserved) victory lap in the popular press. This is the typical pattern in academia, however. What is a good antidote for this? Although I work in another field, I've definitely felt the sting of a bad review for work that didn't conform to the fashions of the day. It is very unjust. Hopefully Hinton and his celebrated cohort will show restraint and give the GOFAI community a break when they are down. I wouldn't be surprised to see that NN(deep learning) will encounter barriers to progress in the future and a variation on GOFAI will prove to be the way forward..

DARPA, the USA sponsor of much AI research is quite balanced, taking risks on many different approaches to challenge problems. Robotics especially, as grounded as it is in physical reality, employs a hierarchy of methods. Chiefly sub-symbolic at lower layers for sensors, e.g. voice recognition and visual object recognition, and symbolic in nature for higher level cognitive functions, e.g. route planning.

I am glad that the AAAI accommodates all the disparate separate research fields that make up AI. With thousands of papers presented, those I enjoyed most aside from the keynotes, were those given by chairs of the specialty conferences describing what's happening in their field. Awesome things are going on - propelled by Moore's law and the contributions from every continent.

I think the article is a tiny bit misleading. The main reason for the perceived lack of enthusiasm towards deep learning in neuroscience stems from the fact that such multilayer networks are not plausible models of neural architecture. In a way they are not unlike chess computers, which surpass human performance multiple times yet offer little to no insight into how a human plays chess. As data analysis tools, however, they are quite brilliant.

Ok, so what are the neuroscientifically, biologically plausible models of how learning and reasoning happen in the brain?

This is astonishingly high quality for pop science. It deserves more upvotes!

Reply on news.ycombinator.com