I'm more in Yann LeCun's camp who called unsupervised learning the cake, supervised learning the icing and reinforcement learning the cherry on top of the cake.
DeepMind says reinforcement learning is ‘enough’ to reach general AI
11–20 of 312 posts
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#12This article is interesting, I even skimmed through their paper. But I think still the question remains: How to find the unified reward function? Or in other words, how to find answer to life? [It cannot be 42].
Yea. For animals, reproduction and just surviving is the reward function?
It talks a lot about having a rich enough environment for learning which makes sense, if a computer lives only in a Go board it can only learn go playing itself.
How do you simulate a rich enough environment purely in software (or do you sense input from the "real" environment) and what reward do we define in this complex environment.. It seems to ask those 2 questions in the discussion but kind of glosses over them imo.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#13If you send a message in a bottle it will eventually land ashore somewhere, maybe in a century, who knows, and who knows whether it will be relevant by then or not, or civilization may not even exist by then, but sure it's similarly plausible to get to AGI via RL.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#14Wait isn’t any Turing complete programming language sufficient to eventually reach general AI
Like the notion that all of natural language can be modeled fully if you just use enough finite state machines — surely true; just a wee bit inefficient.
If you had a custom chip it could be efficient. Determining which state machines would still be hard.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#15Current state-of-the-art in reinforcement learning can barely make a physical robot walk. In theory, with transfer learning, we will probably see better success over time but I'm looking forward to seeing results in practice.
A 2018 article about the challenges of reinforcement learning: https://www.alexirpan.com/2018/02/14/rl-hard.html
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#16Everything old is new again.
As far as I can tell, they're not actually proposing how to achieve this. I can't access the article without a host institution it seems (is there another link?), so I only have the article to go by. RL has been the basis for all robots engaging with the world, and that engagement with the physical world modeled using RL has been promised to make robots that can act like a 2 year old for a long time (see Cynthia Breazeal's work, for example). Yet AFAIK, we haven't actually achieved this as we don't know how to efficiently model the problem to have learning rates that reach anywhere near what we're able to do with DNNs today.
Perhaps someone who has access to the paper can say why this is a milestone? If Patricia Churchland suggests it is, then something new must be happening here.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#17I'm more in Yann LeCun's camp who called unsupervised learning the cake, supervised learning the icing and reinforcement learning the cherry on top of the cake.
My opinion is that view is very simplistic and unnecessarily offensive to a whole class of researchers. MuZero, developed by David Silver, uses a combination of RL, supervised learning, and unsupervised learning (state representation) coupled with a planning algorithm. It accomplished things far beyond anything unsupervised learning can ever accomplish.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#18Some Bozo who has heard all this many times before is suspicious of claims from places like Deep Mind who have a financial incentive to make them (keep funding) where there aren't working machines to back that claim up.
Some Bozo has no credentials, no reputation, no track record of publications and barely supports the claim they're making with anything much. Some Bozo has no financial incentives or otherwise to opine either way. Some Bozo doesn't even work in the field at all.
Bets: Some Bozo or Deep Mind turn out to be closer to being correct in the passing of some finite amount of time? 5 years? 10 Years? 25 Years?
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#19I'm more in Yann LeCun's camp who called unsupervised learning the cake, supervised learning the icing and reinforcement learning the cherry on top of the cake.
Wow, I had no idea that Yann LeCun was also this based. I've figured that it was relatively rare to hear people advocate for the superiority of unsupervised methods - but I guess it isn't if a titan like him does. It's good to hear because epistemologically I just do not believe that most learning is anything but unsupervised. There are very few good labels for our data relative to how much data we process in an information theory sense.
Re: DeepMind says reinforcement learning is ‘enough’ to reach general AI
#20Everything old is new again. As far as I can tell, they're not actually proposing how to achieve this. I can't access the article without a host institution it seems (is there another link?), so I only have the article to go by. RL has been the basis for all robots engaging with the world, and that engagement with the physical world modeled using RL has been promised to make robots that can act like a 2 year old for…
I don't know who Patricia Churchland is, but they said that the paper was "very carefully and insightfully worked out."
After having read the paper, I am very disappointed in the output. Nothing concrete was shown, just hypothesis and reads more like philosophy. That being said, I would say that the paper is carefully worked out and does provide insight if you haven't thought about RL before.