Obstacles to Progress in Deep Learning & AI – Yann LeCun

5 thoughts on “Obstacles to Progress in Deep Learning & AI – Yann LeCun”

1. Great talk ! The 'Learn to Predict Everything' paradigm seems very promising, particularly when combined with Generative Adversarial Networks(GAN). I was reminded of a reinforcement learning paper from Deepmind where the agent learnt to 'imagine' (i.e. predict) the results of its actions. This let it do search to predict the likely reward, a few steps ahead, before choosing an action. The idea worked very well in the completely deterministic domain of Sokoban. However a more probabilistic domain would lead to 'blurry' predictions of future states. It seems to me that GANs are ideal for performing this kind of prediction as they eliminate the blurry outcome. Actually they really do merit the term 'imagine' since they select one possible future from many and compute it in detail. A reinforcement learning program could perform a number of imagination based rollouts, using a GAN, and average over them to estimate the reward.

2. Jeff Hawkin talks a lot about how the brain is basically a prediction machine.

3. "prediction machine"? … don't forget to teach the "problem of induction", so it has some humility! 😉

4. Captions would be great

5. Greetings sir i think you do not have biological AI coding is it .

Comments are closed.