วีดีโอ

When you say, it can train to full accuracy in the same number of steps - isn't that sort of untrue? If it is the same number of steps (N) to train the reduced network as the big network, wouldn't we have 2N at the end of finding and training the winning ticket?

Awesome idea.

Thank you for the insightful presentation

For the initial conditions that work, have anybody look at how much wiggle room you have. Is there an epsilon-neighborhood of the initial state you can safely start from, and how small is epsilon?

Thanks for a great talk

The world of NODE is so small, I can always meet you guys

thanks

Fascinating!

Hi~ Is there a video about ASAM? I wanna learn more about ASAM

Very helpful. A clear explanation of the idea behind the algorithm.

But 0.9 is actually not the only unique answer to the lever problem, there is a unique 1.0 lever directly opposite to the .9 lever. This is stable to the symmetries implied by the problem statement where both players are shown the same set of levers, as well as being robust to different reflections and rotations being presented to different players (though not arbitrary permutations). So assuming whoever I am paired with is also optimizing we would both earn 1. This strategy doesn't work if there are an odd number of buttons to begin with.

what computing power was used to achieve such results ?

I suggest 0.75x speed. Even though my ears can hear Marta clearly but my brain just doesn't catch up with the ideas.

Covers these papers: AWAC: arxiv.org/abs/2006.09359 MOPO: arxiv.org/abs/2005.13239 CQL: arxiv.org/abs/2006.04779 COG: arxiv.org/abs/2010.14500