DeepMind's Demis Hassabis expounds a bit on AlphaGo's innovations, such as the famous move 37 in game 2 in the match against Lee Sedol, and shows a few other examples from later games, and how intuition is implicit knowledge that can't be communicated, but you can test it behaviorally. AlphaGo is more strategically innovative with Go than chess programs have ever been. They did all this with general-purpose learning systems that can be applied in the real world.
17 plus ones
Shared publicly•View activity
- If they had "general purpose learning systems that can be applied in the real world," they would have demonstrated general purpose learning in the real world. They did not. This is demonstration of special purpose learning in a game world.34w
- Your game is done34w
- , in the same video he says they used the same technology to reduce the energy consumption in Google's data centers by 15%. So the learning methods they developed are not limited to the game of Go, or to the game world, but are applicable to the real world. They're working on using them for drug discovery and material design. This is in contrast with program that beat Gary Kasparov at chess which could only play chess.34w
- each application is trained for a rather narrow special purpose. The learning technology itself is in a way more generic than purpose-built special application logic, but each application still needs to be trained for a particular purpose.
Machine learning is not generally applicable to all purposes. In some where you could apply it, writing explicit application logic may be more efficient.
The human process of creating both explicit application logic and other decision automation methods is several orders of magnitude more general than any instance of machine learning, which can be used as a component to handle certain kind of problems. For machine learning to be applicable, the problem needs to be well contained and offer a clear feedback channel or a rather large set of training data discriminating between desirable (correct) and less desirable (false) results. Overall, efficient application of the technology is peculiar to those kinds of problems. Each application itself is peculiar to the chosen kind of a problem it was trained to help solve.
Even if a machine learning solution was better than an unaided human individual in solving a given limited problem, we must understand that the machine is such aid built by multiple humans. Of course a well-funded army of people armed with a fully automatic weapon that they built for a purpose can beat an unarmed individual at that.34w
- That's right, each application is trained for a "narrow" specific purpose -- using underlying algorithms that are are general purpose. The difference between AlphaGo and Deep Blue, the machine that beat Garry Kasparov at chess, is that Deep Blue was written specifically to play chess, while AlphaGo used general-purpose learning algorithms to learn how to play go, primarily by playing millions of games against other versions of itself. Now DeepMind is looking for other applications of the general-purpose learning algorithms they developed.34w
Add a comment...