The Google research blog posted this last week:
It’s stunningly beautiful and gives some intuition as to what is going within the hidden layers of image recognition artificial neural networks (ANNs). It’s not surprising it gained mainstream media coverage. What strikes me as more profound are the implications for psychology. Human minds are after all just massive neural networks, so when we see an ANN doing something interesting, is there a read-across for humans? Well two things jump out at me from this research:
- What the computer scientists have done is something similar to priming in the language of psychology. It basically means you lower the activation levels necessary to trigger a response in a network. So in the Google ANN research, animal heads etc were effectively ‘primed’ to be easily (and in this case erroneously) triggered. The advertising industry spends a lot of money priming you and I to be highly sensitised to certain products. Priming constantly alters our perception and sensitivity to all environmental stimuli. Priming the first thing you learn about in experimental cognitive psychology as an undergrad, and these guys at google have produced an amazing graphic illustration what priming actually looks like.
- The pictures have a hallucinatory, dreamlike quality. One way of changing the activation levels in a real neural network is to mess with neurotransmitters via psychoactive drugs. Looked at in this way, The google researchers seem to have given their AI’s the equivalent of a dose of digital LSD! (or at least that’s the functional outcome, and probably the closest equivalent human phenomena).
What does this all mean? Well, I think we could be on the verge of some real breakthroughs in (human) cognitive neuroscience on the back of current research into deep (machine) learning. I certainly think this kind of research can reveal new ways of thinking about and viewing complex, emergent phenomena in real neural networks.