Thread by @IntuitMachine

Another excellent conversation from the Learning Salon. A lot of participants contributing towards a big picture.

https://twitter.com/neuro_data/status/1317693662139211776

In this episode, @JohnCLangford proposes Reinforcement Learning to be essential to intelligence. An ambiguous statement however since RL isn't precisely defined by him in the talk.

This is opposition to @ylecun icing on the cake analogy. @KordingLab chimed in with an excellent argument against the cake analogy. He insightfully proposes however the immense capability of evolution to absorb information about causality.

@JohnCLangford throws cold water on evolutionary algorithms on the argument of its wastefulness and the pragmatic reality that it requires too much compute resources. It's unclear if he has an overly simplified view of evolution.

John Krakauer @blamlab questions if it is correct to conflate the process that builds something and the functionality of that something. Questions if evolution and RL are even the same thing. Again, a problem of lack of definition of error-correction algorithms.

Krakauer @blamlab further argues the disconnect between something that is learned through gradual methods (equated to procedural) and that learned through declarative methods. The latter is something humans are capable of.

@shengokai argues however that every declarative statement must be grounded in procedural thought. One cannot explain declarative shortcuts on how to ride a motorcycle or swing a sword without prior experience.

@neuro_data
also makes the observation the evolution doesn't have goals unlike RL methods that require goals.
@shengokai
asks what's defining the rewards for an RL algorithm?

So some open questions left that may all be related. How does an error-correcting gradual algorithm lead to propositional cognition? Why does goal-less open-ended algorithms lead to goal based gradual algorithms?

I've updated my big picture: