This is hardly the first attempt to apply neural networks to the game Go, but this result seems noteworthy: "Our convolutional neural networks can consistently defeat the well known Go program GNU Go, indicating it is state of the art among programs that do not use Monte Carlo Tree Search. It is also able to win some games against state of the art Go playing program Fuego while using a fraction of the play time"
Here's a bit on their motivation: "Human Go experts rely heavily on pattern recognition when playing Go. Expert players can gain strong intuitions about what parts of the board will fall under whose control and what are the best moves to consider at a glance, and without needing to mentally simulate possible future positions. This provides a sharp contrast to typical computer Go algorithms, which simulate thousands of possible future positions and make minimal use of pattern recognition. This gives us reason to think that developing pattern recognition algorithms for Go might be the missing element needed to close the performance gap between computers and humans. In particular for Go, pattern recognition systems could provide ways to combat the high branching factor because it might be possible to prune out many possible moves based on patterns in the current position. This would result in a system that analyzes Go in a much more ‘human like’ manner, eliminating most candidate moves based on learned patterns within the board and then ‘thinking ahead’ only for the remaining, more promising moves."