Tongzhou Wang @ssnl_tz Twitter profile | Pikagi

Pikagi

Tongzhou Wang

@ssnl_tz

1,977

Followers

1,141

Following

25

Media

267

Statuses

amortize intelligence, repr learning. trying out new things. ex @mit @pytorch @MetaAI @berkeley_ai

https://t.co/P9taddjE1F

Joined November 2011

Don't wanna be here? Send us removal request.

Pinned Tweet

@ssnl_tz

Tongzhou Wang

5 months

good ML models fit data good AI models find the representation that fits reality 🚀all strong models share _one_ representation, no matter objective/data/modality 👀all datasets are projections of the same reality 🍒representation learning is the goal, not a tool #positionpaper

Tweet media one

Tweet media two

1

12

78

Last Seen Profiles

@tu_deiv1

@Petes_Diary

@CoachMickens

@stakedotlink

@MillMediaCo

@ThyrairfraENKM

@_mahmudpasha_

@Stillpond961

@hwrated

@marce_giano

@ElijahWhosoever

@kelstar_

@TheForexFunder

@nikoniko46498

@ewokakuyochi5

@meire_eloisa

@orioji1980

@irem1849

@dispersa19

@wuntonton

@sotwecom

@WinthropBSB

@NzarAdmn

@bokeplokalmalam

@earthhiyori

@sukautowanchan

@bokeplokalmalam

@dopieropoczatek

@squirrel_7172

@gohenry

@nguyendanxxcn

@nightingale_sh7

@para_classic

@bokeplokalmalam

@diet_yuri_yuri

@tamakinozomu

@ssnl_tz

Tongzhou Wang

11 days

Director & chair of CIFAR, supposedly leader of our field. Yelling at organic changes he didn’t like, blatantly attacking an entire group of AI researchers, as if he dictates the field. Many proper ways to propose change. Screaming “I don’t like it” is childish and is not one.

@roydanroy

Dan Roy

13 days

NeurIPS is now a computer vision conference, it seems. Time to bust it up, because I'm not interested in any of that work, and the top 10 authors are all getting 18+ papers in in areas I don't care about.

30

44

873

12

9

234

@ssnl_tz

Tongzhou Wang

2 years

Goal-reaching / instruction-following is a major driver of new AI advances (LLM as general task solver, controlled generation). The revolution will not be single-task. Our newest work is all about the *geometric* structure emerging in multi-task RL: 1/n

4

53

215

@ssnl_tz

Tongzhou Wang

2 years

I used to think of models as approximations to reality, and the goal is to make them closer and closer to the real world. But these days I'm thinking of models more as improvements on reality. Our paper shows one way that can be true. #ICML Paper: 1/n

Tweet media one

6

28

204

@ssnl_tz

Tongzhou Wang

2 years

It's amazing to see how Dataset Distillation (DD) has become increasingly active & popular in ML. It started from our 2018 paper (, w/ @junyanz89 , A. Torralba, A. Efros) that didn't make it into NeurIPS, ICCV, or ICLR. :) (1/3)

7

25

183

@ssnl_tz

Tongzhou Wang

1 month

AI was divided into too many areas. I'm increasingly convinced that there are only two problems---repr. learning and adaptation.

7

5

183

@ssnl_tz

Tongzhou Wang

3 months

Visa & immigration suck for non-US/EU/… researchers. We are blocked from academic confs. We are blocked from seeing our own family. As much as I appreciate the comforting words and retweets, to make a real difference, pls pressure your org/school to demand the gov to change ❤️

2

10

125

@ssnl_tz

Tongzhou Wang

8 months

Latent space planning. Took me a while to understand why each maze path has two latent paths.

5

7

117

@ssnl_tz

Tongzhou Wang

2 months

The entire field of ML research can easily be 1.5x productive. We just need to have better immigration laws for foreign researchers. The field would probably have 1.5x better mental health too.

1

5

107

@ssnl_tz

Tongzhou Wang

1 year

Quasimetric RL code is now on GitHub: Instead of deleting 80% of the dev repo, I rewrote the algorithm in a hopefully cleaner way. But going through the old repo is fun. So many half-explored interesting ideas in the remaining 80%. RL=geometry

Tweet card media

GitHub - quasimetric-learning/quasimetric-rl: Open source code for paper "Optimal Goal-Reaching...

Open source code for paper "Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning" ICML 2023 - quasimetric-learning/quasimetric-rl

1

23

94

@ssnl_tz

Tongzhou Wang

10 months

I'm @NeurIPSConf this week and on the academic job market. Happy to chat about research (decision making, geometry, repr learning, llm agents), and more!

1

6

67

@ssnl_tz

Tongzhou Wang

1 year

My Hawaii trip ended with running in HNL and barely making to the flight back to Boston At this (likely) my last ICML before graduation, I was blessed with so much incredible nature & amazing people. I hope to see many of you again as I enter the academic job market this fall :)

0

0

58

@ssnl_tz

Tongzhou Wang

1 year

since we are posting Hawai’i activities…. here’s my favorite so far — sunrise hike at Koko crater railway

4

1

47

@ssnl_tz

Tongzhou Wang

1 year

@ hawaii for ICML 20th-29th. I'll present ⬇️ work on RL via learned quasimetric geometry. looking forward to chatting about ML, RL, structures, geometry, abstraction.... or food & activities on the island :)

@ssnl_tz

Tongzhou Wang

2 years

Goal-reaching / instruction-following is a major driver of new AI advances (LLM as general task solver, controlled generation). The revolution will not be single-task. Our newest work is all about the *geometric* structure emerging in multi-task RL: 1/n

4

53

215

0

6

40

@ssnl_tz

Tongzhou Wang

8 months

5 years ago DD couldn't get accepted anywhere. Now there is a workshop at CVPR. It's come along so far.

@GCazenavette

George Cazenavette

8 months

Announcing the 1st Workshop on Dataset Distillation @CVPR 2024! #CVPR2024 More info in 🧵 1/4

Tweet media one

2

29

112

1

2

38

@ssnl_tz

Tongzhou Wang

6 months

Common misconception: Improvement requires new information. Reorganization of existing information is improvement.

6

2

38

@ssnl_tz

Tongzhou Wang

4 years

Excited to share our work on "Diverse Image Generation via Self-Conditioned GANs" CVPR 20. w/ @stevenxliu @davidbau @junyanz89 & Antonio Torralba No labels? No problem! We train conditional GANs with no supervision. 1/2

Tweet card media

Diverse Image Generation via Self-Conditioned GANs

We introduce a simple but effective unsupervised method for generating realistic and diverse images. We train a class-conditional GAN model without using manually annotated class labels. Instead,...

1

7

33

@ssnl_tz

Tongzhou Wang

11 days

@roydanroy Saying something improper, then claiming “LOL twitter”. Reminds me of a certain US president candidate. That’s really a new low. I’m just sad that the field and CIFAR and you tolerate behavior like this. At least I agree w you that twitter is nice allowing me to say what I want.

1

0

30

@ssnl_tz

Tongzhou Wang

2 years

Guang Li, Bo Zhao, and I have started this collection of Dataset Distillation papers. Please check it out and send us PRs! Guang is an explorer of medical applications with DD. Bo is the author of many new DD algorithms (DC, DSA, etc.) (3/3)

Tweet card media

GitHub - Guang000/Awesome-Dataset-Distillation: A curated list of awesome papers on dataset...

A curated list of awesome papers on dataset distillation and related applications. - Guang000/Awesome-Dataset-Distillation

3

6

28

@ssnl_tz

Tongzhou Wang

2 years

I will be at #NeurIPS2022 next week. Excited to meet people and talk about representation geometry, RL, and learning from synthetic data!

4

1

23

@ssnl_tz

Tongzhou Wang

2 years

Updates on learning quasimetrics! 1. A simpler and better method (IQE) @neur_reps : 2. PyTorch package of many quasimetric learning methods:

Tweet card media

GitHub - quasimetric-learning/torch-quasimetric: PyTorch Package For Quasimetric Learning

PyTorch Package For Quasimetric Learning. Contribute to quasimetric-learning/torch-quasimetric development by creating an account on GitHub.

@ssnl_tz

Tongzhou Wang

2 years

Excited to share this work w/ @phillip_isola I'm always interested in how representation encodes data geometry. Here we look at an informative geometric quantity---distance in asymmetrical space, which is more realistic & relevant to many researches (eg goal-reaching RL) 1/3

1

2

14

1

4

23

@ssnl_tz

Tongzhou Wang

2 years

It was great to have Jim @DrJimFan ! The talk was very well done and interesting. If you are interested in flexible environment / large dataset / control, be sure to check it out!

@taochenshh

Tao Chen

2 years

We are excited to have Jim Fan @DrJimFan from Nvidia AI talking about MineDojo today at the MIT CSL seminar. MineDojo is a simulation suite with 1000s of diverse open-ended tasks and an internet-scale knowledge base. Welcome to check out our recording at .

Tweet media one

1

1

36

1

0

23

@ssnl_tz

Tongzhou Wang

10 months

It was truly amazing working on PyTorch in its early days. I learned so much across ML, systems, numerical comp, etc. These learnings still benefit me to this day.

@soumithchintala

Soumith Chintala

@soumithchintala

10 months

PyTorch's design origins, its connection to Lua, its intertwined deep connection to JAX, its symbiotic connection to Chainer The groundwork for PyTorch originally started in early 2016, online, among a band of Torch7's contributors. Torch7 (~2010-2017) These days, we also

43

272

1K

0

0

21

@ssnl_tz

Tongzhou Wang

1 year

Upcoming @NeurIPSConf 23 workshop on Goal-Conditioned RL!

@gcrl_workshop

Goal-Conditioned RL Workshop

1 year

Announcing the NeurIPS 2023 Workshop on Goal-Conditioned Reinforcement Learning (GCRL). We welcome either a 5-minute video or 2-page paper by October 4th, 2023. More info:

1

8

19

0

2

21

@ssnl_tz

Tongzhou Wang

2 years

@icmlconf Author instructions & style files links still point to 2022's.

1

0

19

@ssnl_tz

Tongzhou Wang

6 months

My 2018 hobby proj supports a core part of the modern large-scale training pipeline in fairscale. The true power of #oss

Tweet media one

1

0

17

@ssnl_tz

Tongzhou Wang

8 months

Glad to see that this isn't happening in 2024. Hope that future @icmlconf also takes into consideration LNY and other holidays celebrated by the community. Happy Lunar New Year everyone!

@ssnl_tz

Tongzhou Wang

2 years

I haven't been properly celebrate new year for 3 out of my 4 phd years b/c ICML. Really hope that things can change. It wasn't that bad in 2020 (still not great), but last year and this year's dates are very disappointing.

0

0

12

0

0

14

@ssnl_tz

Tongzhou Wang

2 years

Excited to share this work w/ @phillip_isola I'm always interested in how representation encodes data geometry. Here we look at an informative geometric quantity---distance in asymmetrical space, which is more realistic & relevant to many researches (eg goal-reaching RL) 1/3

@phillip_isola

Phillip Isola

2 years

New work at ICLR this week: “On the Learning and Learnability of Quasimetrics” Project page: Paper: w/ @TongzhouWang 1/n

1

5

37

1

2

14

@ssnl_tz

Tongzhou Wang

2 years

The real world presents us information so rich that processing every detail is impossible. Intelligence requires identifying and extracting useful signals. E.g., humans plan with abstracted mental models that exclude irrelevant factors. Artificial agents should do the same. 2/n

1

0

14

@ssnl_tz

Tongzhou Wang

2 years

This continues our line of work on quasimetrics: + PQE @iclr_conf 2022: Learnability of quasimetrics via different models + IQE @neur_reps 2022: SOTA quasimetric model + torchqmet (): Your go-to package for *quasimetric models*. 12/n

Tweet card media

GitHub - quasimetric-learning/torch-quasimetric: PyTorch Package For Quasimetric Learning

PyTorch Package For Quasimetric Learning. Contribute to quasimetric-learning/torch-quasimetric development by creating an account on GitHub.

1

2

13

@ssnl_tz

Tongzhou Wang

2 years

DD is the task of *synthesizing* a *small* dataset s.t. models trained on it perform well on the original large real dataset. A good small distilled dataset can be useful in dataset understanding, continual learning, privacy, NAS, etc. (2/3)

1

1

12

@ssnl_tz

Tongzhou Wang

10 months

Come check out our workshop tomorrow and the great line of speakers!

@gcrl_workshop

Goal-Conditioned RL Workshop

10 months

Check out the #NeurIPS2023 workshop on Goal Conditioned Reinforcement Learning: * Tomorrow (Friday) 900 -- 1830 CST * Great speakers: @jeffclune @r_mirsky @olexandr @ybisk @SusanMurphylab1 * Program:

0

2

17

0

1

12

@ssnl_tz

Tongzhou Wang

2 years

Can we directly learn V*? Q: Two obj's are connected by many chains. How to find length of the shortest chain connecting them? A: Push them apart, and their distance in space will be limited by the shortest chain. Measuring this distance gives the length of shortest chain. 5/n

1

0

12

@ssnl_tz

Tongzhou Wang

2 years

I haven't been properly celebrate new year for 3 out of my 4 phd years b/c ICML. Really hope that things can change. It wasn't that bad in 2020 (still not great), but last year and this year's dates are very disappointing.

@zdhnarsil

Dinghuai Zhang 张鼎怀

2 years

Same thing happens to ICML this year (and serveral ICMLs in the ast years) that the deadline is only a few days before Chinese New Year’s Eve. Imagine one has to stay stressed out during Christmas holiday… and this happens to authors of one third of the ICML submissions.

1

4

47

0

0

12

@ssnl_tz

Tongzhou Wang

10 days

@SashaMTL @JeffDean @GaryMarcus @bate5a55 @strubell you can't do that when this number is affecting how a large group of people think about this issue. evidence: how this number was organically brought up as reference in this thread

0

0

11

@ssnl_tz

Tongzhou Wang

2 years

@hciprof But this is also misleading. COVID spread is not uniformly random. It is easier in close social circles. So a close group of people tend to have either (1) higher than avg or (2) lower than avg infection rate. Neither stat shows the complete picture.

0

0

11

@ssnl_tz

Tongzhou Wang

3 months

If you have had a non-US/EU/… collaborator, or even simply enjoyed a paper from them, I urge you to think about how their struggle relates to your field, and to your career.

0

0

11

@ssnl_tz

Tongzhou Wang

2 years

Paper: "Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning" arXiv: Project: This work would be impossible without my collaborators A. Torralba, @phillp_isola @yayitsamyzhang . 11/n

2

1

10

@ssnl_tz

Tongzhou Wang

6 months

@jon_barron It is from NCHW

Tweet card media

Add F.normalize by szagoruyko · Pull Request #1467 · pytorch/pytorch

with gradcheck and doc.

0

0

9

@ssnl_tz

Tongzhou Wang

1 month

@phillip_isola Another recent realization is that sys1/perception is in fact repr learning, rather than classification etc.

0

0

9

@ssnl_tz

Tongzhou Wang

8 months

@rsalakhu @PhilBeaudoin @ylecun The society (and academia) rewards confidence. E.g., your tweet will be less popular if it starts with "I think I maybe agree with ..."

0

0

8

@ssnl_tz

Tongzhou Wang

2 years

Essentially, we have showed various benefits from a learned model that represents the real world as factorized (signal, noise). Denoised MDP Paper (ICML 2022): Project: Code: 7/n

Tweet card media

GitHub - facebookresearch/denoised_mdp: Open source code for paper "Denoised MDPs: Learning World...

Open source code for paper "Denoised MDPs: Learning World Models Better Than the World Itself" - GitHub - facebookresearch/denoised_mdp: Open source code for paper "Deno...

1

1

8

@ssnl_tz

Tongzhou Wang

5 months

Every now and then a game strikes me as an epitome of human creativity, design & execution: Braid, SM: Odyssey, Tunic, and just now, Animal Well. How far are we from an AI that can build such games? I’m sure execution is solvable. Is the rest 0, 1, N, or ∞ breakthroughs away?

0

0

8

@ssnl_tz

Tongzhou Wang

2 years

@neur_reps I am really excited about the package: Check it out!

Tweet card media

GitHub - quasimetric-learning/torch-quasimetric: PyTorch Package For Quasimetric Learning

PyTorch Package For Quasimetric Learning. Contribute to quasimetric-learning/torch-quasimetric development by creating an account on GitHub.

0

1

8

@ssnl_tz

Tongzhou Wang

11 days

@tdietterich @roydanroy It certainly is. And the right question is how can we provide better venue choices for both CV and ML theory. And the right tune is NOT claiming CV is the problem bc its “low standards” eats away the papers “I am interested in”.

0

0

8

@ssnl_tz

Tongzhou Wang

5 months

@francoisfleuret ask and you shall receive gist:

Tweet media one

1

0

11

@ssnl_tz

Tongzhou Wang

2 years

There are many useful ways to represent the world! See our other newly arXiv'ed paper on how quasimetric geometry relates latent spaces to optimal control. Quasimetric Paper (ICLR 2022): Project: Code: 8/n

Tweet card media

GitHub - ssnl/poisson_quasimetric_embedding: Open source code for paper "On the Learning and...

Open source code for paper "On the Learning and Learnability of Quasimetrics". - ssnl/poisson_quasimetric_embedding

2

1

7

@ssnl_tz

Tongzhou Wang

8 months

+ Policy trained to follow latent geodesic (yes not strictly planning). + Embedding points colored based on latent distance to goal (≈ env dist). + Dim reduction via UMAP on symmetrized latent distance.

3

0

7

@ssnl_tz

Tongzhou Wang

5 months

⬆️ is what we posit in #ICML24 position paper: The Platonic Representation Hypothesis * why & how models share one repr.? * what are we converging to? * relation to scaling/AGI/multimodal/arch.? * X is not converging? * lossy projection? see paper & 👇🧵

@phillip_isola

Phillip Isola

5 months

New paper: The Platonic Representation Hypothesis In which we posit that _different_ foundation models are converging to the _same_ representation of reality. paper: website: code: 1/8

35

256

1K

1

1

6

@ssnl_tz

Tongzhou Wang

2 years

Denoised MDP, our online model-based algorithm, uses factorized transitions to automatically discover such information, and extracts them in the form of a *compressed* latent world model. Any planning / policy training needs only to be performed w.r.t. this compressed model. 4/n

Tweet media one

1

0

6

@ssnl_tz

Tongzhou Wang

1 year

@pcastr @g_sokar @agarwl_ @utkuevci Nice paper and video, but when’s the song version of the summary coming out?

2

0

6

@ssnl_tz

Tongzhou Wang

2 years

The potential of such learned models/representations is encouraging and leads to many interesting questions. Are there other useful structures? How to combine them? We are are excited about this direction and looking forward to sharing more as we make more progress. n=10

0

0

6

@ssnl_tz

Tongzhou Wang

4 months

@adad8m Isn’t this the ellipse that is tangent with the circle with A and B being the focal points? We can just find the diameter that equally divide line seg AB.

0

0

4

@ssnl_tz

Tongzhou Wang

2 years

Denoised MDP is able to cleanly learn and separate signal from various noise distractors across many environments built on DMControl and RoboDesk. In the video, see how only information relevant to the task is changing in the Signal column. 5/n

1

1

6

@ssnl_tz

Tongzhou Wang

8 months

Cool paper on how to train online multi-env actor-critic with a shared tsfmer backbone. I wonder what representation it learns.

@yukez

Yuke Zhu

8 months

One of RL's most future-proof ideas is that adaptation is just a memory problem in disguise. Simple in theory, scaling is hard! Our #ICLR2024 spotlight work AMAGO shows the path to training long-context Transformer models with pure RL. Open-source here:

Tweet media one

1

34

190

0

1

6

@ssnl_tz

Tongzhou Wang

2 years

@AnthropicAI Nice post! But I'm afraid this is incorrect "Scaling a Gaussian produces a Gaussian" since the scale applied depends on the value of the Gaussian. Similarly, if x~N(0,1), x/|x| is a scaled Gaussian but is definitely is not a Gaussian.

2

1

5

@ssnl_tz

Tongzhou Wang

2 years

PQE repo () has a simple API and fast CUDA kernels Going forward, I'm super excited to see what good latent geometric priors can achieve, especially in data-limited settings (eg RL/robotics) We'll be at #ICLR2022 poster session 3 tomorrow (4/25)! 3/3

0

1

5

@ssnl_tz

Tongzhou Wang

6 months

@francoisfleuret (1) the reference code is wrong. above_diag should use <. cumsum should be sum. (2) cumsum on B gives what you want

Tweet media one

2

0

5

@ssnl_tz

Tongzhou Wang

5 months

@alexgraveley The less lossy the projections are, the more aligned the repr (fig 9). IMO lossiness is exactly what makes things interesting in practice - How to use other modalities to recover the lost info - How learning should focus on info-dense data - How to even measure info density etc

2

0

5

@ssnl_tz

Tongzhou Wang

3 years

@svlevine The added term is identical to the KL term in Dreamer and SLAC. Dreamer even derived from the same MI. While optimizing pi wrt it is new, I think its form should be better attributed to earlier work. The method is also a model-based SAC, but comparison with SLAC is missing.

1

0

4

@ssnl_tz

Tongzhou Wang

8 months

@EugeneVinitsky Short horizon or tabular are great stable settings for explaining basics. It’s also easy to see how things can become unstable when they are not satisfied.

1

0

4

@ssnl_tz

Tongzhou Wang

2 years

@EugeneVinitsky @ptrblck_de

0

0

5

@ssnl_tz

Tongzhou Wang

2 years

In MDPs, V*(s; goal) is the *optimal* costs between two states, and satisfies *triangle-inequality*. Optimal cost A→C ≤ Optimal cost A→B→C. Dynamics are generally irreversible ⇒ V* is a *quasimetric*, a generalization of metric that doesn't require symmetry. 2/n

Tweet media one

Tweet media two

1

1

4

@ssnl_tz

Tongzhou Wang

5 months

this paper connects many topics at the center of my research: representation, generalist agents, learning, AI, generalization w/ @minyoung_huh @thisismyhat @phillip_isola (all equal contrib.) see paper for more findings and maybe controversial arguments:

1

0

4

@ssnl_tz

Tongzhou Wang

9 months

@francoisfleuret Yeah. Here the recurrent state would have to encode Union(useful_info_for_all_futures). The potential issue I see is when there are too many possible futures, it may be hard. When temp=0, there is a fixed future.

0

0

2

@ssnl_tz

Tongzhou Wang

6 months

@ZimingLiu11 @2prime_PKU Why is activation on edges/nodes a big difference? Tied weights make them interchangeable.

1

0

4

@ssnl_tz

Tongzhou Wang

8 months

how latent geodesic ⇒ optimal plan Left embeds s → [ d_z(s0, s), d_z(s, g) ]

0

0

4

@ssnl_tz

Tongzhou Wang

2 years

Our new RL algorithm, Quasimetric RL (QRL), works by the same principle! QRL uses *quasimetric models* for V*, ensuring *triangle-inequality* + *ability to approx. any MDP*. It pushes random states apart, *subject to* a constraint that local transition costs are preserved. 7/n

Tweet media one

Tweet media two

1

0

4

@ssnl_tz

Tongzhou Wang

11 days

@JoeyZhouUT I also consider myself working in both ML and CV. I believe that subfields should respect and work with each other to solve such issues, rather than pointing fingers and blaming the other community as being the root of the bad things.

1

0

4

@ssnl_tz

Tongzhou Wang

6 months

I see nothing wrong with obtaining effectively new information by applying known skills/fns on better organized known knowledge, and internalize and repeat indefinitely.

0

0

4

@ssnl_tz

Tongzhou Wang

5 months

@minyoung_huh @thisismyhat @phillip_isola very fun collab. would do it again

0

0

4

@ssnl_tz

Tongzhou Wang

2 years

@EdwardRaffML @BlackHC @CsabaSzepesvari @bremen79 Then we should ask: why is it normal in academia for people to change their plans for next 2 days (for abstract) and 5 days (for paper), based on a Saturday decision?

0

0

4

@ssnl_tz

Tongzhou Wang

2 years

Two things made this answer work: (1) *Triangle-inequality* of the Euclidean physical space (2) Each link of the chains has a *fixed length unaffected by our pushing* 6/n

1

0

4

@ssnl_tz

Tongzhou Wang

7 months

@EugeneVinitsky "gen ai is killing students' ability to learn" is probably what people are actually concerned about

1

0

2

@ssnl_tz

Tongzhou Wang

6 months

fairscale for FSDP: my proj for reparam neural nets:

Tweet card media

GitHub - ssnl/PyTorch-Reparam-Module: Reparameterize your PyTorch modules

Reparameterize your PyTorch modules. Contribute to ssnl/PyTorch-Reparam-Module development by creating an account on GitHub.

1

0

3

@ssnl_tz

Tongzhou Wang

2 years

We categorize information in MDP observations based on controllability and relevance to reward, where optimal control only relies on information both controllable and reward-relevant. 3/n

1

0

3

@ssnl_tz

Tongzhou Wang

5 months

@chris_j_paxton not-good-enough features is the limiting factor

0

0

2

@ssnl_tz

Tongzhou Wang

8 months

@sainingxie the blog's so great!

0

0

3

@ssnl_tz

Tongzhou Wang

2 years

@sucholutsky @junyanz89 So many cool ideas in recent papers! The most exciting thing is that they keep coming and don't seem to slow down!

0

0

3

@ssnl_tz

Tongzhou Wang

11 days

@MihaiCNica He didn’t care if broader communities agree with him. He just makes claims solely based on his own beliefs w/o verifying them. He thinks change should happen just because “I” am not interested in CV works, and “I” believe CV has low standards and made “my” experience worse.

1

0

3

@ssnl_tz

Tongzhou Wang

11 months

@seohong_park very cool paper :)

1

0

2

@ssnl_tz

Tongzhou Wang

1 year

@xuanalogue @jhsansom Yessss let's treat different things in its own proper way: constraints, cost, task success, preference, etc. They (& combinations of them) should not be considered just a scalar fn, that is optimized with the same alg.

1

1

3

@ssnl_tz

Tongzhou Wang

5 months

@TacoCohen @roydanroy The probability theory has a very well defined P with meaurable spaces and events. IIUC you should really complain about people not using the shorthand in a clear way, rather than about prob theory.

0

0

3

@ssnl_tz

Tongzhou Wang

1 year

@nanjiang_cs @yoavgo @unsorsodicorda @matteohessel @s_scardapane @carperai The argument is that “say idk when the model doesn’t know” behavior is hard to learn in SL, and that the model can’t know everything.

1

0

3

@ssnl_tz

Tongzhou Wang

1 year

@daibond_alpha @neu_rips @nanjiang_cs @braham_snyder @francoisfleuret (Shameless plug) Here’s a value-based alg. that 1. uses a form of LP characterization of V/Q* 2. learns V/Q* without TD 3. is empirically strong (4. assumes determinism)

@ssnl_tz

Tongzhou Wang

2 years

Goal-reaching / instruction-following is a major driver of new AI advances (LLM as general task solver, controlled generation). The revolution will not be single-task. Our newest work is all about the *geometric* structure emerging in multi-task RL: 1/n

4

53

215

1

0

3

@ssnl_tz

Tongzhou Wang

6 months

@ElronBandel @naaclmeeting nice work! but I think you should really cite prototypical net

Tweet card media

Prototypical Networks for Few-shot Learning

We propose prototypical networks for the problem of few-shot classification, where a classifier must generalize to new classes not seen in the training set, given only a small number of examples...

0

0

3

@ssnl_tz

Tongzhou Wang

6 months

@xuanalogue Then I hope people will be incentivized to make sure all their submissions are good papers.

0

0

3

@ssnl_tz

Tongzhou Wang

2 years

@EugeneVinitsky @MichaelD1729 @xuanalogue I'd love to know the answers too. I expect a goal curricula that gradually expands the frontier would help. Also, maybe the estimated distance here can be used to create the frontier (it is a consistent heuristic after all)? Lots of interesting questions!

0

0

3

@ssnl_tz

Tongzhou Wang

1 year

@tydsh About writing, too. Whether the phrasing helps reach the intended audience, and makes it easy for them to understand.

0

0

3

@ssnl_tz

Tongzhou Wang

2 years

@phillip_isola This paper: 1. Why quasimetric learning 2. Empirical failure of common methods 3. A large family of methods _provably_ fail (proved w/ probabilistic method!) 4. Our method (called PQE) has good guarantee and performance 5. Preliminary results on social graph and Q-learning 2/3

1

0

3

@ssnl_tz

Tongzhou Wang

6 months

There should be an easy, automated process to get official certificates for reviewing X papers at Y conference.

0

0

3

@ssnl_tz

Tongzhou Wang

2 months

@EugeneVinitsky I think we should do a better job exercising. The benefit-to-effort ratio is much higher than most research activities.

1

0

3

@ssnl_tz

Tongzhou Wang

2 years

It has been truly wonderful experiences working with many awesome people on these projects. I enjoyed and learned a lot from these collaborations. Denoised MDP w/ @SimonShaoleiDu Antonio Torralba @phillip_isola @yayitsamyzhang @tydsh Quasimetric Learning w/ @phillip_isola 9/n

2

0

3

@ssnl_tz

Tongzhou Wang

2 years

We also would like to shout out to Contrastive RL by @ben_eysenbach and MRN by @cranialxix , both of which influenced our work. n=13

0

0

3

@ssnl_tz

Tongzhou Wang

5 months

@cHHillee @francoisfleuret i learned this from your flops tracker xD

0

0

2

@ssnl_tz

Tongzhou Wang

2 years

Better denoised models give better representation, improving policy learning as well non-control tasks such as robot joint position regression. 6/n

Tweet media one

Tweet media two

1

0

2

@ssnl_tz

Tongzhou Wang

8 months

@xuanalogue reasoning = inference local relations -[learning]-> global inference

0

0

2

@ssnl_tz

Tongzhou Wang

11 months

@jon_barron @ChrisWu6080 @BenMildenhall @philipphenzler @KeunhongP @RuiqiGao @watson_nn @_pratul_ @dorverbin @poolio @holynski_ learned prior + strict geometry structure from data ❤️

0

0

1

@ssnl_tz

Tongzhou Wang

2 years

@xuanalogue @MichaelD1729 Our paper made this assumption clear, and noted that it is shared by many recent RL works. In practice, QRL still works well in some of our experiments with stochastic dyn.

0

0

2

@ssnl_tz

Tongzhou Wang

2 years

There is a wide range of interesting discoveries to be made about QRL (repr. learning, exploration, etc.)! We are ever more excited about quasimetrics and other structures in general multi-task learning! 10/n

1

0

2