Artificial Intelligence – Zero to hero.

Reply to this post

DeepMind AlphaZero is smarter with 1000x less effort.

• DeepMind has taken another step forward in the quest for machine intelligence with the demonstration of the rapid training of a single algorithm to play Chess, Go and Shogi.
• While this is without doubt another step forward, I do not consider that the second major challenge in AI is close to being solved.
• RFM has identified three main challenges that need to be overcome for AI to really come of age (see here).
• These problems are:
  • First: the ability to train AIs using much less data than today,
  • Second: the creation of an AI that can take what it has learned from one task and apply it to another and
  • Third: the creation of AI that can build its own models rather than relying on humans to do it.
• DeepMind’s previous publication took a shot at problem one (see here) and while it represented an advance, I did not consider it to have really solved the problem.
• Its current publication (see here) takes a shot at problem two, but again has made an advance, but in my opinion, has not really cracked the problem.
• DeepMind describes a new algorithm called AlphaZero which is a generic version of AlphaGo Zero (Go algorithm (see here)).
• It uses a deep neural network instead of the specific policy and value neural networks that were designed to play Go in AlphaGo Zero.
• AlphaZero is then given the rules of Chess, Shogi (Japanese version of Chess) and Go and asked to play itself and to use reinforcement learning to improve.
• In each case AlphaZero was quickly able to obtain a level of play that allowed it to beat the best algorithm available in each of the three games including the original AlphaGo Zero.
• It is also highly relevant that AlphaZero did far less “thinking” than its opponents.
• Each machine was given 1 minute of thinking time and during that time AlphaZero searched 80,000 positions per second for chess and 40,000 per second for Shogi while Stockfish (Chess) searched 70 million per second and Elmo (Shogi) searched 35 million per second.
• In effect, AlphaZero expended 1000x fewer resources to arrive at a better solution than its opponents due to its use of its deep neural network to tell it where to search.
• The ramifications for this are substantial as it implies that once trained, algorithms could be easily and efficiently executed on mobile devices where resources remain extremely constrained.
• However, it is critical to recognise that for each game, DeepMind trained a different instance of AlphaZero.
• DeepMind started with three instances of AlphaZero which were all identical other than they each had the rules for a different game.
• However, through playing themselves and reinforcement learning they all diverged from one another as they gained expertise in the specific game they had been asked to play.
• The end result is that despite a common starting point, the three algorithms become very different by the time that they are capable of playing these games at a very high level.
• Consequently, to me this does not represent the solution to problem two because one cannot take the Chess version of AlphaZero and have it win at Shogi.
• However, what it does do is represent a major step forward in the training of algorithms as the AlphaZeros all trained themselves and they all came from a common starting point.
• This should make training of algorithms in the future easier, quicker and cheaper than they are today which is why this is yet another very significant advance that has been made by DeepMind.
• Seeing that DeepMind is owned by Google, it is Google Ecosystem devices and services that are likely to benefit from these advances long before anyone does.
• This will allow Google to differentiate its services more effectively and make them more appealing to users.
• We gave already seen signs of this where Google is able to do portrait mode with one camera when everyone requires two.
• This reconfirms my position that it is Google that leads the world in AI developments for digital ecosystems with Baidu and Yandex in 2^nd and 3^rd
• Given, Alphabet’s exceptional stock performance this year, Baidu now makes the most interesting and cost-effective entry point for anyone looking to gain exposure to AI.