What makes deep neural networks so effective?

Four lessons down in the course, I’m fascinated by the fact that deep networks can be applied in a wide array of domains with such a high efficacy. What makes them work so well? I did some reading and came across papers that try explaining why deep networks work so well. The basic idea is that deep networks work better than shallow networks because our (physical) world is “deep” too.

I converted my learnings on this topic into a 16 minute Youtube video. Here it is: https://www.youtube.com/watch?v=Y-WgVcWQYs4

Hope you like it! This is my second time making a video, so if you have feedback for me, it’ll help me improve.

There’s a question posed at the end of the video (starting at 14:38): if deep networks work by encoding structure of real world into its weights and biases, is it possible to reverse engineering deep networks to discover new knowledge is possible? Are you aware of any attempts of doing it?

(@jeremy, I hope I’m not spamming by posting my own video. I’ll remove this post if it’s not appropriate)