Google DeepMind has reported substantial progress on one of the big three challenges of AI, which is exactly what Facebook desperately needs but is unlikely to achieve anytime soon.
DeepMind has been able to build a new Go algorithm (AlphaGo Zero) that relies solely on self-play to improve, and within 36 hours was able to defeat AlphaGo Lee (the one that beat Lee Sedol) 100 games to 0.
RFM has identified three main challenges that need to be overcome for AI to really come of age:
- The ability to train AIs using much less data than today.
- The creation of an AI that can take what it has learned from one task and apply it to another.
- The creation of AI that can build its own models rather than relying on humans to do it.
In my opinion, DeepMind’s achievement represents a huge step forward in addressing the first challenge, as AlphaGo Zero used no data at all. I do not think that this represents a step forward against the third challenge, as the system of board assessment and move prediction (but not the experience) used in AlphaGo Lee was also built into AlphaGo Zero.
Hence, I do not think that this system was building its own models, but was instead using a framework that had already been developed to play and applying reinforcement learning to improve.
What will really have the likes of Elon Musk quaking in their boots is the fact that AlphaGo Zero was able to obtain a level of expertise of Go that has never been achieved by a human mind (see here figure 3). It is almost as if the use human data limited the potential of the machine’s ability to maximize its potential.
That being said, it is one thing to become superhuman at Go and quite another to enslave the human race. So I continue to think that dystopia will continue to be thwarted by Dr. Moore.
There have been many other attempts to address the data quantity problem but this is the first one that I have seen that has shown real progress. Many of the other digital ecosystems have been trying to use computer generated images to train image and video recognition algorithms, but there has been no real success to date.
I suspect that taking what DeepMind has achieved and applying it to real world AI problems like image and video recognition will be very difficult. This is because the Go problem is based on highly structured data in a clearly defined environment whereas images, video, text, speech and so on are completely unstructured. Hence, we are not about to see a sudden improvement in Google’s ability to recognize and categorize images and video (which is already world-leading), but the seeds are clearly being sown that will keep Google a long way ahead of everyone else.
This exactly the kind of advance that Facebook really needs to make. This is because I have long been of the opinion that while Facebook sits on a massive treasure trove of data, it has very little idea of what any of it is or what it means. This makes it very hard to spot fake news or offensive content which has been the source of many of Facebook’s most recent problems. It also makes it much more difficult to understand what its users do and do not like and therefore much more challenging to tailor its service accordingly.
Finally, it will also make it much more difficult for Facebook to keep up with competition in terms of deep and rich services – meaning that its users may begin to spend time elsewhere.
This is a breakthrough that Facebook badly needs, but unfortunately it is Google that owns the IP, meaning that it will be Google services that improve.
This article was originally published at RadioFreeMobile