Here’s to throwing open the AI gates….
The issue of Nature magazine published on October 19, 2017 had an interesting, and surprisingly low-key paper about a self-learning Artificial Intelligence (AI) algorithm (read here a nice summary if you can’t get into the complete paper right now).
With virtually every magazine, blog or newspaper worth its tech-salt carrying at least one Machine-Learning (ML)/AI/algorithms article per issue these days, this one didn’t seem like a big deal then and I skipped it.
I came across the paper again only recently. This time, as I read through it, I realized how important this development really is. To understand the level of disruption this paper talks about, we must realize two things. One is obvious that success or failure of an AI/ML initiative is vastly dependent on its training. Second: training, in turn, is dependent on quality and quantity of data that you use in building it.
Many know the example of a machine that was taught to recognize dumbbells (you know- the bar-with-two-rounds – used in gyms by people other than me). After training however, this machine could not recognize photo of dumbbells taken in a gym. It had been trained on a set of images that almost always showcased humans lifting the dumbbells. So, the machine couldn’t recognize a dumbbell that came without a human hand attached to it.
This example is as much about choosing the right set of data as it is about sheer quantity – the more you have, the less likelihood of missing something. That’s why most of AI/ML applications (not research) today come out of companies that already own tons of data – tech giants like Google, Apple, Microsoft, Facebook, Amazon and more. It’s been commonly assumed that these companies will continue to lead fields of AI/ML in future.
Well, what start-up could compete with say, a Facebook, which even tracks your mouse hovers today?
However, this paper in Nature outlines how this could possibly change.
Published by DeepMind which is ironically owned by Google now, the paper speaks about a version of their famous AlphaGo gaming algorithm called ‘AlphaGo Zero’. Unlike earlier versions however, the latest version of this algorithm was not trained on data sets of previous games – it trained itself on the game of Go by playing and understanding the rules. In just three days of training, this algorithm beat Alpha Go Lee – the previous version of Alpha Go that famously beat Lee Sedol in 4 out of 5 games back in 2016 (here’s.a very interesting account of that)
Equipped with just basic rules of the game, AlphaGo Zero started playing itself over and over. In 19 hours, AlphaGo Zero had started to play complex strategies and in 70 hours, it started coming up with unconventional strategies not used by human players before. In three days it beat AlphaGo Lee in 100 out of 100 games.
And all of this was done after being taught just basic rules of the game.
Of course, it’s all very specific to playing the game of Go – self-learning AI may not be possible in areas where there are no clearly-defined inalienable rules in like, say, brain tumor detection. However, something like this could democratize AI/ML development by not needing access to massive clean data sets – at least in applications that come with clear rules.
This could be in beginning of an era where we’ll create AI algorithms to train other AI algorithms.