Google On AlphaGo’s Win Over World Go Champion: “We Landed it on the Moon” – How Impressive is the Win and What Does it Really Mean for the Future of AI?

By Babak Hodjat, Co-founder and Chief Scientist

Google’s AlphaGo just defeated the world champion Go player Lee Sedol 3-0 in what can only be described as a landslide victory. Was it to be expected? Has AI really come this far and do displays of machine intelligence like these really have an impact on solving real world problems? Let’s review case-by-case.

25717574115_a13441f167_b

In 1997, Garry Kasparov, the reigning world champion chess player, was beaten by IBM’s Deep Blue, heralding what people thought of as a momentous time in the history of AI. The Deep Blue chess computer that defeated Kasparov in 1997 could search to a depth of twenty or even more moves in some situations. Aided by its massively parallel, 30-node RISC-based UNIX servers, Big Blue had encoded 4,000 positions and 700,000 grandmaster games and could evaluate 200 million positions per second. According to IBM, the goal behind Deep Blue was to show intelligence that could solve the “kinds of complex calculations needed to help discover new medical drugs; do the broad financial modeling needed to identify trends and do risk analysis; handle large database searches; and perform massive calculations needed in many fields of science.”

This was a significant win, not just because the computer had overpowered a human in what used to be our exclusive domain of prowess, namely intelligence, but because it underscored the power of game-playing strategies and human-engineered heuristics running at scale using what was, at the time, state-of-the-art supercomputing hardware.

The Big Blue machine was later dismantled by IBM but reincarnated as Watson in 2011 when it went on to beat champion Jeopardy winners in a nail-biting two day event. The machine was so big it had to be kept backstage, so the audience could not hear the noise of the fans required to keep the 2,880 Power7 cores cool. This achievement showed AI had moved on to master natural language and showcased the sheer speed and accuracy of knowledge retrieval by regularly beating its rival contestants to the buzzer.

It was successful in part thanks to its vast compute capacity and its ability to try hundreds of different algorithms, each trying different approaches to understanding the clue.  IBM has since started to productize its IBM Watson platform with some degree of success.

Then we turn to Google’s AlphaGo win. It is a remarkable achievement by Demis Hassabis and his team at DeepMind. The number of possible games that can be played with Go is of a magnitude far larger than that of chess → 10 to the power of 47 compared to 10 to the power of 170 for Go → making the win much more impressive than the one against Kasparov. Given the complexity of the Go problem, merely using scaled strategies to exhaustively search the tree of possible moves and outcomes, the approach taken by Deep Blue in its games,  is not possible, despite all the recent advances in hardware and supercomputing.


Enter Machine Learning.

This milestone in AI marks a coming of age for Machine Learning, demonstrating how this technology, at scale, can learn to defeat the best biological learners amongst us.

DeepMind trained its machine on a sample of 30 million Go positions culled from online servers where amateurs and professionals gather to play. The crunching of this data is typically done on data centers of connected computers, powered by thousands of CPUs, each augmented by thousands of smaller, specialized GPUs. The resulting models, driven by patterns and strategies of play identified and constructed by the algorithm off of these samples, can easily be run on a single machine.

While this achievement is immensely significant, there are limitations. The current state of the art in Machine Learning driving AlphaGo requires considerable manual engineering by humans in order to configure the learner for best results on the task at hand, in this case, playing Go.

The resulting system, while quite impressive, is very specialized. The models, for instance, from playing Go, cannot generalize to playing another board game, say chess, unless the system is reset, reconfigured, and re-trained specifically on chess moves. Something that takes time and domain expertise.

Finally, and perhaps most significantly, the learning power of AlphaGo is bound by a single data center with high bandwidth interconnections in amongst its computers. The monolithic nature of the neural networks being trained means that the integrity of the solutions need to be maintained throughout training. In other words, if one of the machines in the data center finds a beneficial tweak to the model, all other machines need to become aware of the modified model immediately. There are physical limitations to how large we can build such data centers.

The good news is that Distributed AI aims to tackle these very limitations, and important strides are being made to solve them. At Sentient, for example, we have already broken the barrier of being restricted to a single highly interconnected data center, scaling our algorithms to millions of CPUs in a robust manner, geographically dispersed and sporadically available. This scale will allow us to all but eliminate the need for manual configuration of learners, allowing the system itself to find the best configurations. These breakthroughs drive the successes in our trading and visual intelligence products.

Machine Learning is already being used in many real-world applications, solving complex problems no longer within reach of humans. I don’t know what the next grand challenge will be, but I am certain it can only be overcome by massively scaled distributed AI.