AlphaGo Google DeepMind

AlphaGo Google DeepMind surpasses human intelligence in one of the world’s most complicated board games. What’s more, is that its victory came against one of the best players of our generation. The implications here are huge-in both the professional and personal sphere; what this means for human competitiveness and employment rates is still unclear.

The research team behind AlphaGo, a system designed by researchers at Google’s DeepMind subsidiary to play Go, has revealed how it managed to beat South Korean grandmaster Lee Sedol 4:1 in a five-game series last week (March 9th). It was not only the first time artificial intelligence (AI) program has beaten a top player without handicapping itself, but also something no human being has ever accomplished. That may sound somewhat hollow considering that DeepMind’s AlphaGo program is just a computer algorithm, but the implications are huge.

source: google

This achievement now places technology at least one step ahead of humans in an area that requires intuition and an understanding of long-term strategy to prevail over your opponent. It also means that the program can be brought to other similar “perfect information” games where skill is the only determining factor, such as Chess or Shogi.

Training for this match began about two years ago when AlphaGo was still little more than a collection of what are known as deep neural networks—vast networks of hardware and software that approximate the web of neurons in the human brain. Like their biological inspiration, these networks are capable of learning through repetition and adjusting to feedback based on games played between other programs or against themselves.

The training process was split into two parts. The first used a pair of neural networks that looked at the possible outcomes from each move to predict which would lead to the highest probability of victory. This is known as a policy network and was trained using data from 30 million moves by the original AlphaGo program over the past two years.

The second, known as reinforcement learning, then played games between copies of itself or against itself while trying to improve its game in relation to the earlier neural network prediction. This second set of games led to incremental increases in AlphaGo’s overall performance, with the ultimate result being that it was able to beat the original program 100-0 when they played each other. The same process was then used for training against Sedol.

By combining two different neural networks in this way, DeepMind believes it has created a more general-purpose algorithm that can be used for other purposes. “We think that we have invented a new paradigm for learning,” Demis Hassabis, the founder and CEO of DeepMind told WIRED before the match.

This is just the latest advancement in artificial intelligence technology from Deepmind which has been making strides in gaining an advantage over humans recently.

In January, the company unveiled a system that could be used to discern finer details from MRI scans of a patient’s brain. In February it revealed an AI program that was capable of learning games on its own-and mastering tasks with no prior information. Earlier this month, Deepmind also taught AI how to master arcade games by learning trial-and-error.

The immediate applications here are still unclear, but it’s possible to make some educated guesses about what the future may hold. One likely possibility is that this technology could be used to replace lower-level employees with AI systems in an organization.