Re: The future of deep learning ...

Another interesting IEEE Spectrum article:

 https://spectrum.ieee.org/how-deepmind-is-reinventing-the-robot <https://spectrum.ieee.org/how-deepmind-is-reinventing-the-robot>

This talks about the use of simulated worlds as way to bootstrap robot learning in the real world. Unfortunately, “simulated worlds are too perfect, too removed from the complexities of the real world. You can add noise and randomness artificially, but no contemporary simulation is good enough to truly recreate even a small slice of reality.”  That said, simulated worlds can be used to evaluate how well robots can learn, so that they can learn effectively once exposed to the real world.

The article also talks about the problem of catastrophic forgetting, where when an AI agent learns a new task, it tends to forget all the old ones. This is motivating work on mitigations, e.g. knowledge distillation and elastic weight consolidation, which combines an active network for learning new skills, and a stable network to preserve older skills. Other approaches are inspired by how we dream, reactivating neurons in similar patterns to those that arose when it was having the corresponding experience. “In between learning tasks, the neural network recreates patterns of connections and weights, loosely mimicking the awake-sleep cycle of human neural activity. The technique has proven quite effective at avoiding catastrophic forgetting.” 

I am wondering whether such ideas could help with mimicking how we learn language. Learning new words shouldn’t result in unlearning older ones. I can also envisage using simulated worlds as a way for cognitive agents to interact with one another, analogous the children in the playground.

Dave Raggett <dsr@w3.org> http://www.w3.org/People/Raggett
W3C Data Activity Lead & W3C champion for the Web of things