Single Iteration Learning

Behold, Mice:

The first one is the learning round, the 2nd is the “learned” round. When it knows what to do. It goes like the clappers. I especially like the way it does a fancy little 1/2 spin when it’s finished.

I think these mouse-races are vaguely interesting – because they demonstrate something… that as usual, I can’t quite put my finger on… something to do with the difference between bits and atoms.

I’m drifting (or at least attempting to drift) from programming to micro-industry. From bits to atoms… and one of the things that is really striking is the shift in headspace required. With bits, once you learn something it’s done. “To Name Something Is To Have Done With It”. Once you learn something, you can abstract it away… “It Is Known” – which means you never have to think about it again.

Don’t work so well with atoms. You can nearly do it… with the CNC end of things – which is (in case you were wondering) to do with the purity and uniformity and “unnaturalness” of the materials that are used… but I’ve been making those golden mean calipers for 2.5 years now, and I’m still learning techniques etc. One of the revolutionary tectonic undercurrents of the current age is the drive to turn hardware problems into software problems. That is what CNC (in it’s myriad forms) does. Software is easier than hardware – in part because of the abstraction thing. Once something is learned, it is known, and can be given a name and “invoked” but never thought of again. Software has a bias towards single-iteration learning.

I think there might also be some relevance to AI – as in machine intelligence that learns as it goes – from feedback that it gets from somewhere. My new Niece is a type of AI… a couple of months old… and it’ll take years of full-time training to get her to work properly. There’s a slightly older model – a couple of years… which is a vast improvement on the new one, but which still has soooo many years before she becomes fully-functional, that it’s bewildering. It’s a process of massive multi-iteration learning. A huge programming job.

So anyway, that new mouse is learning how to do something that requires a single iteration. Normal reality is not so accommodating – and is so complicated and unreliable that it will probably take an AI to negotiate it… or a mix of hard-coding and AI (which is what we meat-bots tend to have). The the alternative is controlling then environment (the art of politics) – which is what lego is, and laser-cut perspex is and what this maze is and so on. Existing industrial robots.

Maybe that’s another huge tectonic undercurrent of the robotics revolution – creating systems with reduced needs for controlled environments. Until that becomes the norm, then the creation of controlled environments/inputs is probably going to be at least as big as the creation of the machines that will operate within them.