We are neural nets

Back in the late '90s, Penrose's Emperor's New Mind approach to how we work rather appealed to me. That had brains working as a kind of quantum machine, putting a certain amount of specialness in the way they work, and rather nicely tying free will to quantum randomness.

Since then, the last decade of artificial intelligence work has frankly astonished me. Neural net programs have been around since at least the '50s and, well, around 2000 they were more than a little unimpressive. Progress was rubbish, artificial intelligence consisted of everything we didn't know how to do, and so couldn't make computers do.

After that, advances in deep learning (TM) have done some fantastic things. We have image recognition, with textual description generation and go-playing programs able to beat top professionals, to name just two applications of neural nets that would have been mind-boggling just a few years ago. They are also applications in which the systems can demonstrate some very human-like behaviour.

If you look at the Deep Dream images, you can see some more human-like traits - it's spotting patterns that aren't really there, and emphasising it, producing psychodelic images - i.e. images reminiscent of how a misfunctioning vision system works, but also, well... rather imaginative.

For a while I thought that, even if quantum effects weren't key to how our brains work, there might be some "secret sauce" in the biomechanical systems, that the implementation of our brains had some extra subtlety. Looking at what we can achieve by building relatively simple neural network models, it seems that the actual detailed hardware of how our brains work is just an implementation detail, of as much relevance to the overall algorithms our brain runs as the CMOS process used is relevant to a processor's ISA.

A lot of of our high-level models of the brain are fitted neatly by just saying "we're a big neural net". "Priming" and related behaviour can be viewed as just activating the subnetworks associated with that concept. The idea of thinking as dealing with a network of semantically-related symbols (e.g. Goedel, Escher, Bach) is somewhat more literally true, albeit by throwing lots of nodes at the problem, rather than trying to have an explicit node for each concept.

Sleep looks suspiciously like running our brains in "training mode". "Deep Dream" is perhaps a particularly good title, in that it shows the effects of reinforcing the patterns being spotted. We are, at a very fundamental level, pattern recognition machines, and this explains why we like picking up new patterns so much - why we like to learn new things.

If we really are so closely related to the artificial intelligences we are building today, it starts to make the terminology itself suspect. If we keep going, the intelligence we produce won't be fake in any way, as "artificial" may imply (although it will still keep the meaning of deliberately constructed). Perhaps "machine intelligence" is better.

On the other hand, if we are, effectively, the implementation of a learning algorithm, the way we think may have some universal roots. If we ever do make contact with aliens, we may find them more like ourselves than we expected.

Posted 2016-03-13.