Twitter | Search | |
hardmaru
research scientist at google brain 🧠 香港人 🏴 ⠀
17,215
Tweets
680
Following
69,265
Followers
Tweets
hardmaru 11m
Replying to @dh7net @NeurIPSConf
Thanks! Let’s chat at the creativity workshop
Reply Retweet Like
hardmaru 7h
Come by our poster this afternoon if you want discuss learning world models that are not forward models! 05:30—07:30 PM East Exhibition Hall B + C #188
Reply Retweet Like
hardmaru 8h
Deep Equilibrium Models Their method is equivalent to running an infinite depth (weight-tied) feedforward net, but has the notable advantage that they can analytically backpropagate through the equilibrium point using implicit differentiation. Cool work!
Reply Retweet Like
hardmaru 8h
Replying to @sarahookr
Good crowd around the poster
Reply Retweet Like
hardmaru 15h
Replying to @rupspace
Love your avatar 🙃
Reply Retweet Like
hardmaru Dec 9
Replying to @hardmaru
What’s a more interesting research area for you? Would you rather your spend time trying to understand mouse-level intelligence, or would you rather try to figure out this small gap between humans and mice?
Reply Retweet Like
hardmaru Dec 9
Humans are more intelligent than mice, but only by a small tiny margin. But this small tiny epsilon amount makes the world of difference.
Reply Retweet Like
hardmaru Dec 9
Humans are not that much better ...
Reply Retweet Like
hardmaru retweeted
/MachineLearning Dec 9
Reinforcement Learning Upside Down: Don't Predict Rewards -- Just Map Them to Actions
Reply Retweet Like
hardmaru Dec 9
Replying to @rupspace
Companion paper by et al., with some experiments.
Reply Retweet Like
hardmaru Dec 9
Reinforcement Learning Upside Down: Don't Predict Rewards—Just Map Them to Actions “This Imitate-Imitator concept may actually explain why biological evolution has resulted in parents who imitate the babbling of their babies.”
Reply Retweet Like
hardmaru retweeted
hal 👾 Dec 9
GitHub runs one of the most important bits of digital infrastructure on the planet, the world’s most popular open-source code repository. Now it is exploring an expansion in China, one which may include operating under Chinese law
Reply Retweet Like
hardmaru retweeted
Hong Kong - Be Water Dec 9
Sheung Wan MTR Station on Dec 8, a MTR staff tried to stop a "Doraemon" from entering the paid area. Other passengers pointed out that "Doraemon" was not causing inconvenience to anyone and he could easily take the lift instead of the escalator.
Reply Retweet Like
hardmaru Dec 9
But other locations have better food
Reply Retweet Like
hardmaru Dec 9
Replying to @EmtiyazKhan
Reply Retweet Like
hardmaru retweeted
Thomas Lahore Dec 8
This is a SUPER interesting idea! Training Agents using Upside-Down Reinforcement Learning "in upside-down RL ... the roles of actions and returns are switched."
Reply Retweet Like
hardmaru Dec 8
Training Agents using Upside-Down Reinforcement Learning RL algorithms either predict rewards with value functions or maximize them using policy search. We study an alternative: Upside-Down RL, that solves RL problems primarily using supervised learning.
Reply Retweet Like
hardmaru Dec 8
Replying to @mitra6ce @jpmorgan
Congrats!
Reply Retweet Like
hardmaru Dec 8
Replying to @jpmorgan
Hmm, claimed that this paper had been accepted, but I couldn't find it in the list of accepted conference papers, or at any workshop. Maybe I wasn't looking hard enough.
Reply Retweet Like
hardmaru Dec 8
Reinforcement Learning for Market Making in a Multi-agent Dealer Market They build a multi-agent simulation of a dealer market to study market behavior. RL agents can be trained manage inventory and learn to skew prices against counterparties.
Reply Retweet Like