Search the Community
Showing results for tags 'Reinforcement Learning'.
Found 1 result
For all you AI and Singularity geeks out there, I found this article entitled "How to Build a Mind?" really interesting and intuitive. It is about a new theory that attempts to integrate recent findings from both neuroscience and deep learning. It explains how the hippocampus encodes recent episodic memories of life events, and then plays them back while we sleep in order to "train" the neocortex. This is very similar to the approach used by the folks at DeepMind to teach AlphaGo using a combination of supervised learning and reinforcement learning. In fact, Demis Hassabis, founder of DeepMind, was a co-author on the paper, along with neuroscientist Jay McClelland from Stanford - a former colleague and collaborator of mine during the earlier days of artificial neural networks while we were both at CMU, along with neural net pioneer Geoff Hinton, who is also now at Google. In fact, most of the smartest people I've ever worked with (literally, at least 10 people I can think of off the top of my head) now work at Google... Which reminds me, earlier today I watched a really good video (embedded below) by Demis about DeepMind, AlphaGo and the future of AI. Demis describes in very accessible detail (starting at 28:15) the way AlphaGo works, and how it was trained (by playing against itself millions of times) to beat Lee Se-Dol, one of the world's top human Go players. I predict that DeepMind's approach to the development of artificial general intelligence may actually work, and come to fruition in the next couple decades. That will really make things interesting. I usually shy away from offering investment advice, but for those of you who haven't but can afford to, I recommend investing at least a little of your retirement savings in Google as an insurance policy against technological unemployment. When and if a Google AI steals all the jobs, you'll be glad you did... --Dean https://www.youtube.com/watch?v=f71RwCksAmI