Today’s post features teaching artificial intelligence as well as laser chaos. I know the combination of the two sounds like a recipe for disaster, but please hear me out before you dive into your nearest bunker.
Lasers are interesting not just because they’re lasers, but because they have really interesting properties. Thanks to electrons hopping around from the lasing material (if you’re thinking light-saber type stuff, you’re not too far off), laser ‘beams’ are really just what happens when electrons have to let go of energy. Unlike your everyday flashlight, which releases all the colors (or at least a certain color range depending on what the glowing material is made of), lasers release a very narrow band of colors, maybe even just one very specific wavelength (like infrared lasers, not unlike what’s in a remote control). That’s where we get the idea of ‘laser-like focus’ from, pretty much – one specific color, one specific point.
Turns out that with a little know-how, engineering types can get lasers to end up ‘chaotic‘ in that the electrons that I mentioned earlier aren’t at the energy levels that they should be in – there’s a long story here that’s far, far away from casual but mostly it means that the laser beam is just a bit more random. This is something that’s gotten the attention of computer scientists in the late 20th century (we can actually say those last three words!). See, these lovely men and women harness the power of chaos itself – they may have other terms for this like randomness or entropy – to, among other things, make our computers more secure. They harness the power of chaos for cryptography.
Computers can generate random numbers, and gamers are very familiar with this concept as part of a gaming mechanic (although most call it ‘luck‘). However, the random numbers that a computer can generate are likely to come from some equation. It can be a highly complicated one, sure, but an equation nonetheless – that means one can, with a lot of brain power or an efficient computer program, find patterns and crack the code. That’s why it’s ‘pseudo-random’ rather than truly random. Where can you find ‘true randomness’? According to random.org (I already put the link), you’re likely to find it in nature.
This week’s Featured Article (now I get to it!?) applies laser chaos to solving really complicated math problems, specifically the multi-armed bandit problem. Yes, there is a mathematical routine for picking the best slot machine in a row of slot machines in Las Vegas. Mathematicians must have a thing for random numbers and gambling. Wouldn’t be the first time this connection’s been made.
It must’ve been successful, since it’s up for publishing. You don’t often to get to publish if nothing happens (but it can be done). Anyway, they got the laser to go all chaotic by reflecting a small fraction of the beam back into the laser device itself; sci-fi tropes tell us that this is where the bad guy’s laser device explodes, but we’re dealing with milliwatts of power at about 295 K. That’s about as powerful as a couple of laser pointers and about about as hot as a pleasant summer morning. The chaotic signal is fed into a machine that selects what slot machine to play (remember, it’s the many-armed bandit problem I mentioned earlier) a total of 400,000 times. It only took their computer 1.2 seconds to run the whole thing – not bad for 16 GB of RAM. I’ve got a computer with 16 GB of RAM and it’s very good at taking apart my club in Football Manager 2015 within 15 minutes into the first half. I’ve disappointed myself as an Arsenal fan…
Naruse et al. declared it a success as over time (in nanoseconds, a small fraction of the time it takes to blink), the program started picking the correct machine at every play, even after the authors switched some conditions around. That is incredibly fast.
So we can use lasers to teach AI to make decisions at ludicrous speed. So what? Well, speed is great and all, especially when we’re dealing with teaching self-driving cars, but what stood out to me was the low overhead involved. These aren’t very powerful lasers, the working temperature is comfortable, the computer isn’t all that sophisticated – embedding something like this in a car probably wouldn’t be all that expensive, so it’s learning while it runs. Anyone who’s ever seen the end of WarGames know that split-second decision-making and learning could be critical when deciding upon best courses of action. For now, at least we know that Watson could really make bank in Las Vegas if we give it some lasers.
I feel like at this point I should bring up something about AI ethics, but I’ll leave that to people who have done the hard thinking and have experience with robotics. Some already have. For now, I’m making verbal and physical threats against the AI lurking in my computer, because Nottingham Forest put three past Arsenal, I’m out of the FA Cup, and the board wants to talk to me. My career as a blogger-manager-scientist is about to be missing a word…
Featured Article: Naruse M, Terashima Y, Uchida A, Kim SJ. (2017) Ultrafast photonic reinforcement learning based on laser chaos. arXiv preprint arXiv:1704.04379
Featured Image: Public Domain (Author: LoggaWiggler, 2014)