What an utterly interesting time to be alive. We all remember how Deep Blue defeated Kasparov. Chess became a ‘solved problem’ pretty soon after that historic event. And in the past week, we have witnessed an even more amazing feat: AlphaGo beating world-class Go player Lee Sedol. The game that knows more positions than there are atoms in the universe was no match for Google’s DeepMind team.
It’s fascinating to see how a game with relatively simple rules can lead to such complex and strategic gameplay. It reminded me of Conway’s game of life. A simple world with simple and deterministic rules yet it harbours so much emergent behavior.
What’s behind all this behavior? How does changing the rules and initial state influence the possible outcomes? Can we learn something from this relation? If such a clean, tidy world already gives rise too this much complexity, it’s no wonder our chaotic world with its opaque rules leads to even more unpredictable outcomes. Still, we seem to be able to predict some of those outcomes with increasing accuracy.
Deep learning, and machine learning in general, are making big waves in the industry. From self-driving cars to self-cleaning inboxes, neural networks seem to be everywhere. As is the overly optimistic conclusion that Artificial General Intelligence (AGI) this time surely looms on the horizon. While AlphaGo’s success is impressive, the superlatives uttered on for example Slashdot (“We know now that we don’t need any big new breakthroughs to get to true AI”) don’t do justice to reality.
What we see is a general technique being successfully applied in many different areas, with some fairly hefty domain-specific algorithmic tuning for each domain. Feeding it a huge amount of existing training data then seals the deal. That’s not enough to get to general intelligence in machines. Many advances in unsupervised learning are necessary before we can get even close. And to be honest, I find that to be a comforting thought. While it’s inevitable that machines chip away at our humanity piece by piece, I’m rather happy with the current status quo. Let us control the machines for a while longer…
In reaction to AlphaGo’s dominance, South Korea announced a whopping $863 million dollar fund for Artificial Intelligence research this week.
“.. thanks to the ‘AlphaGo shock’, we have learned the importance of AI before it is too late”
So it’s not just Silicon Valley anymore chasing after AGI. I believe more countries will follow South Korea’s lead and invest heavily in AI research. Much like the space race in the 20th century, or the nuclear arms race. There is so much potential power for those who are on the forefront. Maybe we’ll even see proliferation treaties on AI knowledge. Imagine what a rogue nation (or corporation) might want to use this technology for.
Back to earth
The applications of machine learning are endless and adoption is in its infancy, notwithstanding outliers like AmaGooFaceSoft. How do we, as Luminis, get ahead in this area? Sure, we can enrol in machine learning MOOCs like so many others do. I know many of you have, myself included. It’s a great start. Your enthusiasm increases, and you’ll at least appear to know what you’re talking about.
But doing machine learning projects is qualitatively different from ‘traditional’ software development. This already starts with (potential) customer interaction. Before starting with software development, you really need to get a feel for what data the customer has and how it can be leveraged. Then, implementing this data-driven insight may not even need that much custom code when building on top of existing machine learning platforms. It does, however, require an intimate knowledge of learning algorithms, statistics and the application domain.
I really believe there are many companies out there who do not know yet what potential there is in their own data. Let’s get them into this century, while improving ourselves at the same time!
Time for the music. To start your week with some energy, here’s Game of Life by Circus Maximus. I promise I’ll select a more laidback track next time…