In this talk, Demis Hassabis describes how the best approach to Artificial General Intelligence probably involves understanding and implementing the brain's algorithms. It affirms most of Numenta's values, and I found it to be an interesting listen. He even mentions the HTM briefly.
Demis Hassabis: Combining systems neuroscience and machine learning: a new approach to AGI http://vimeo.com/m/17513841 I was especially fascinated by his description of the neocortical / hippocampal system at the end. He says that the hippocampus implements short-term episodal memory, and replays these memories at high speed to the neocortex during sleep. It chooses memories to replay stochastically, while giving higher weighting to more emotionally salient memories. Thus it learns a more conceptual and relevant representation of the world rather than being limited by the statistics of the world's model. I wonder, how much of this is proven? He also talks about reinforcement learning, and how neurons that activate on dopamine release associated with the presence of a light cue learn to activate earlier, anticipating the light cue. It reminded me of the temporal pooling idea. I would love to learn more about details on what we know about how the brain implements reinforcement learning.
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
