you're always in a box. being aware of the box can help you tremendously. it's when you think that you've left the box that's dangerous, because you're still in the box, but now you don't know it.
You're always in a box, but it's not necessarily a cubic box, and it usually has more odd nooks and crannies than you'd expect. Most people restrict themselves to a certain well-explored subset of your problem space. The trick is to explore the rest of it.
of course exploring the box from the inside can never tell us if the box is bright purple and unusually shaped, because there is no outside to stand in.
if aliens ever come along, they might tell us things that are along the lines of "well of course your box is purple how could you not see that?" and of course we'll be able to tell them that their box is lime green.
People should stop associating neural networks with backpropagation networks. That is like saying all of AI is based on simplistic rule-based systems. I guess it is mostly people with only casual knowledge of both fields who make suggestions of that sort.
Also, there is no reason to get outside the box. Most research and VC money tends to concentrate in one corner of the box. Go forth, be brave, and explore the box first!
I don't object to "backpropagation" aka gradient descent, which is a bit of simple calculus-based optimization. What I object to is the widespread public perception that saying "neural networks" is a brilliant new paradigm-overthrowing Key to General Intelligence, after over three damn decades.
Biologically inspired stuff? I applaud it to the extent that it works. It's not a magic key to anything, and it doesn't avert the challenge of understanding.
it also seems the surest path to producing an intelligence that acts as if it were a product of natural selection, i.e. one that is selfish, xenophobic, etc. (other sexual competition based heuristics).
By "simplistic rule...", I implied something deterministic. If you define a rule as something that can include stochasticity, then how is a neural network something more than a stochastic graph with rules for communication?
Considering that the backpropagation and variants thereof gained significant commercial/industrial adoption while "monkey and banana" type problem solving systems mostly failed, I think his criticism of artificial neural networks is quite unwarranted.
The entire change has been that people aren't trying to go for 'general intelligence'. They approach it purely as a statistics problem. We have some data, what can we do with the data, how accurately can we do that?
