If you want to really understand how contemporary neural-net-based AI1 works, and you have 18 minutes to spare, you should watch the video linked at the end of this post.
Geoff is unapologetically dogmatic in this presentation; he takes strong positions that are not (yet) universally accepted, though most are strongly supported by observational evidence. Full disclosure: my own intellectual framework is entirely aligned with what he presents.
What is especially beautiful about the model presented is that it offers explanatory power in both the digital and biological domains and, moreover, a powerful bridge between them.
This is a high-information-density talk; for example, around the 9:30 mark, Hinton frames semantic (re)construction as an exercise in self-assembly. Beautiful. There is a lifetime’s worth of research questions in just this one remark.
Enjoy2.
And if you fully believe, as I do, in what Hinton presents, then this will also help you really understand how you think and represent knowledge.
Or be enraged, if you are a Chomksy-loving linguist or AI “neat”.