Reward Functions
This is a fascinating talk on deep reinforcement learning. In my opinion it is a little bit click baity. I think it is a far cry labeling these types of systems as anything close to AGI. Even the environments that these agents learn in are contrived and can arguably never model reality. We can get extremely close to simulating physics but arguably we can't get to that 100%. Now for some problems, like walking, a high fidelity physics model might not be needed. I think for AGI though, a very high fidelity physics simulation would be needed. That's because there is evidence that even quantum effects play a role in cognition. And so simulating cognition may require a low level physics simulation going all the way down to quantum level and below. That's a very tall order to fulfill.
I disagree with the sentiment that we are close to AGI as he suggests. With that said I think that deep learning is a new tool in the toolkit. It's a powerful computational paradigm, but it is not the holy grail, AGI.
Maybe someone should apply these reinforcement learning methods to the OpenWorm project, because AFAIK they are still relying on a hosh posh ensemble of dozens of different models to achieve their result. Once someone figures out a way to simulate the c. elegans in high fidelity, than I think we can possibly make the argument for AGI. Personally, I think deep learning is fundamentally different from the approach that'd be needed to achieve AGI. What that unique approach is, I don't know.
It's also not enough that the same algorithm is used to achieve good results across a number of different problems. Also, network topology is usually determined by experts and it takes domain specific knowledge to come up with the ideal topology.
Ultimately, everything that he showed is a sub-problem. Solving many sub-problems using solutions engineered for specific sub-problems is a lot less impressive than solving many or all sub-problems using the same solution. Reinforcement learning and deep learning in general has achieved incredible results across a large number of sub-problems, but each solution has to be engineered for the specific sub-problem.
Just merely encoding the problem input data in a manner that can be used with a DNN requires domain specific expert knowledge. Just the simple problems he demonstrated in the video have vastly different input encodings and solutions. I think deep learning is amazing, but I think AGI will have to be fundamentally different.
I also don't think the belief that the top-down symbolic approach of AI will meet the bottom up DNN approach in the middle and that's how AGI will be achieved. This just seems like a cop out.
This video makes compelling arguments for why dnn's can't be used for end-to-end self-driving. I think this has some implications about the limitations of dnn's.