Sergey Nivens/Bigstock.com

Hinton after AlphaGo reflects on Big Numbers

Among the various commentaries on the AlphaGo victory earlier this month, a real coup was registered by Canada’s Maclean’s Magazine in securing an interview with Geoffrey Hinton. Generally regarded as the godfather of deep learning, Hinton is especially well-placed to see the AlphaGo victory as an advance for neural networks — adaptive learning systems that mimic the working of the human brain in spotting patterns, intuiting rules, and projecting conclusions far beyond the powers of standard computer programming.

Once established, a neural network effectively practices with and “thinks” for itself. It’s rather like the baby that seems to be lying there, absently playing with its toes: it’s actually working out how language works. Once the Deep Mind programmers had established AlphaGo’s grasp of the rules and strategies of the ancient board game, it could be left to its infinite workings out of the trillions of move permutations. It is said that when GO sat down with human champion Lee Seedol, it had played over 30 million games, mostly with itself. The Korean is estimated to have played some few thousand games in his 33 years. It is amazing he won the one game out of the five.

 

But Hinton goes deeper into the numbers and, along the way, shows how neural networks of both the human and the machine variety are wondrous phenomena in their different ways. Evolution has enabled our brains to develop a much vaster universe of connections, or synapses, within their neural networks of cells – outnumbering the machine equivalent by a factor of a million. And the AlphaGo’s computations would have consumed hundreds of kilowatts of power, as against Sedol’s getting by on 30 watts, just about what it takes to power a lightbulb.

So the potential in consolidating the best of both types of neural networks into one super intelligent thinking system hints at a future far beyond our current imaginings.

Lively Stuff from Planet BAM!

Leave a Reply

Your email address will not be published. Required fields are marked *