The Bitterest of Lessons: The Role of Data and Optimization in Emergence

A talk by Sergey Levine about emergence

[JimW]

This is a really interesting and well explained lecture. Thank you for uploading.

[haluk]

18:17 it is nice of you to accept the fact that RL is the remnant of psychology, behaviorism (animal learning) and their disfigured cognitivist child adopted by CS. The cognitivist child is love child of "between perception and action there is a soul, we cannot study the soul but we can study its input output" has married with "there is nothing called soul, we dropped the thing in BETWEEN" and there you got "there has to be something in BETWEEN, we call it cognition". hence the false action-cognition-perception loop.

[haluk]

15:57 your search definition is acceptable, but "to search in order to satisfy a need optimally" is the complete picture. Without a need, you don't "need" to search. It is already embedded to language, when you don't need, you don't. Hence now comes the need model. you need to have a model of the world that is "good" and "bad", which are both about consciousness, not an Q-function.

[haluk]

16:53 data and optimization doesn't mean anything without a need to satisfy. ı know many things I can do and I only do one thing and not the other. because I compare those searches that is possible with the evidences I was faced up before.

[haluk]

14:58 no, learning is interacting the thing in focus and creating generative models with the help of the evidence. Patterns and likelihood aren't the same, data and evidence aren't the same. You don't interact with the data. And patterns are not necessarily causal.

[haluk]

41:54 "Discovery without theory" is oxymoron, because by definition it is impossible.

[haluk]

19:38 offline? we have a saying in our language "you do not enter the nuptial night with someone else's genitals".

[haluk]

4:10 wrong, you see to act, but you act to see, as well. In fact, most of the time you act to see. You have a need model "when to stop and satisfy", and you act on it, in a way "model minus perception" drives action most of the time. Literally "what is not perceived among the possible perceivable futures" is the driver of actions and it drive perceptions. if you don't want to include paradoxes in your paradigm, you have to assume both action and perception is due to the model, and action & perception is integrated along their the way.

[haluk]

18:41 ah yes, being a cute baby actually gives you enough times and caretaker and safety and cirriculum to make the thing on the right. I wonder if there is a trend called "Baby AI" that wants to tackle this problem?

[haluk]

17:08 by the way, decision is control, so they are the same.

[haluk]

2:36 constraint optimization is not emergence. limiting something quantitatively has nothing to do with the solution's emergence.

[haluk]

23:51 yes, but unless you mapped the goodness and badness of that experience and ALL possible experiences by hand (or by your own grounded cognition and memories, wink) you cannot extract the best, because if the data is the true thing to mimic, data is never wrong and everything the computer has done is right. If there are ways to be mistaken, you have to attribute valence to ALL (seen or not seen).

[haluk]

2:00 Optimization is not emergence. It is free energy doing work to get down to a local minima. It is the most physical physics ever, not emergence.

[haluk]

1:27 your emergence definition is broken beyond fixing. "What we didn't expect". There is no way to measure your expectance and your causal relations in your head, what you didn't expect may be a totally expectable thing to someone else. You cannot just define emergence as "trick a human being" in the most Turing Test style and expect a progress. It is not emergence because you didn't expect, you may be having a bad week and lowered your expectations all together.