Our
#ICML2024
@icmlconf
paper "Diffusion Model-Augmented Behavioral Cloning" led by Shang-Fu Chen and Hsiang-Chun Wang augments BC by learning a diffusion model to capture expert demos and provides gradients to train the policy.
Paper:
A thread 👉
I am teaching the very first Reinforcement Learning course at National Taiwan University this semester! Dear RL folks, can you describe why students should learn RL in one sentence? Please RT 🙏
Review
#1
: Weak Reject
Review
#2
: Weak Reject
Review
#3
: Reject
No one:
AC: Ok. I read the reviews and the rebuttal. All the concerns are addressed. ACCEPT.
This is an AC that only exists in your wildest dreams.
#ICLR2020
I am beyond excited to share that I'll be joining the Electrical Engineering Dept. at National Taiwan University (NTU) as a tenure-track Assistant Professor this fall! I am thrilled for this next chapter and am extremely grateful to everyone who has been part of my Ph.D. journey.
#NeurIPS2023
A reviewer just wrote “thanks for you explanation” to our 600-word rebuttal addressing all the questions he/she asked. Exactly these four words. Just that simple. I’m in shock. Who am I speaking to?
I defended my Ph.D. thesis! Words can't describe how thankful I am to my family, advisor
@JosephLim_AI
, collaborators, dissertation committee members (
@snikolaidis19
,
@gauravsukhatme
, Prof. Quan Nguyen), and those who joined the defense today.
I am proud to present my very first last-author paper as an Assistant Professor at NTU. I'm incredibly fortunate to work with these dedicated and talented students and colleagues: Guan-Ting Liu, En-Pei Hu, Pu-Jen Cheng,
@HungyiLee2
A thread 👉
With a paper accepted by NeurIPS, I finally completed my first machine learning "grand slam": publishing papers at NeurIPS, ICML, and ICLR as a leading author. I am so grateful to all my collaborators as well as the researchers who discuss ideas with me and provide feedback.
1/n This
#ICML
#ICCV
plagiarism issue reminds me of an uncomfortable experience of mine. I believe that the code of my ICLR 2020 paper Program Guided Agent (), which is not publicly released
I am thrilled to announce that our NeurIPS 2019 Spotlight paper ”Multimodal Model-Agnostic Meta-Learning via Task-Aware Modulation” is now on arXiv ()!
#NeurIPS2024
Among the 24 reviewers for our six submissions, 18 responded to our rebuttal (had back and forth discussions with 6), and 13 raised their scores. This is the most engaging rebuttal I have ever experienced. Kudos to our reviewers and ACs!
I just passed the qualifying exam.
I am now officially a Ph.D. candidate!
Special thanks to my qual committee members
@JosephLim_AI
,
@snikolaidis19
,
@yanliu_usc
,
@gauravsukhatme
, Prof. Rahul Jain for the valuable feedback and questions to my research!
#ICLR2020
submissions have an average score of 3.9067. If your paper got an average score of 5, it beats 81.79% of submissions. Hot keywords are RL, representation learning, generative models, graph neural network, etc.
Our
#ICLR2020
paper "Program Guided Agent" is accepted as a spotlight paper! We propose a modular framework that can accomplish tasks specified by programs, structured in a formal language, and achieve zero-shot generalization to more complex tasks. Paper:
If you are interested in how programs can play a role in improving the generalization of machine learning methods, check out this list of papers related to program synthesis, program induction, program execution, program repair, and programmatic RL.
I crowdsourced a one-sentence motivation for learning RL on Twitter to kick off the RL course I am teaching at National Taiwan University. I appreciate all the replies! Here's a thread 🧵 organizing the answers I got.
I am teaching the very first Reinforcement Learning course at National Taiwan University this semester! Dear RL folks, can you describe why students should learn RL in one sentence? Please RT 🙏
#ICLR2024
@iclr_conf
score statistics (7304 papers):
mean: 5.10; max: 8.67; min: 1.00
>8.5: top 0.59-0.67%
8.0: top 0.68-1.75%
7.5: top 1.93-3.35%
7.0: top 5.20-8.17%
6.75: top 8.42-10.34%
6.5: top 12.13-15.39%
6.25: top 17.48-20.56%
6.0: top 21.02-28.29%
5.75: top 29.07-33.27%
As asked by
@danijarh
and others
#ICLR2020
authors with >= 5 submissions sorted by acceptance rate
Zhiyuan Li, Mingyuan Zhou, Deva Ramanan 4/5 80.0%
Le Song 7/9 77.8%
Jimmy Ba 6/8 75.0%
Martin Jaggi 5/7 71.4%
Abhinav Gupta 5/7 71.4%
Pushmeet Kohli 6/9 66.7%
Max Welling 5/8 62.5%
The first paper from my lab at National Taiwan University w/ my students Hsiang-Chun Wang and Shang-Fu Chen. We employ diffusion models to augment behavioral cloning (BC) and achieve competitive performance in various continuous control domains.
A thread👉
#ICLR2024
Just arrived in Vienna, Austria 🇦🇹! I will be at ICLR from Monday to Saturday. Drop me a line if you would like to chat! Also, any tourist spots or restaurant recommendations are appreciated!
I just booked my
#NeurIPS2024
@NeurIPSConf
hotel through the official hotel reservation portal. Staying at Pan Pacific Vancouver (closest to the venue) costs $560+/night on Agoda, hotels, etc., and it's only $190 on the portal! Kudos to the organizers.
#ICLR2024
@iclr_conf
pre/post-rebuttal scores
Pre-rebuttal: mean 4.87; max 8.50; min 1.00
Post-rebuttal: mean 5.10; max 8.67; min 1.00
The average score increase this year is probably the lowest compared to all previous ICLRs I have experienced (2019-2023). Why?
I just released our
#NeurIPS2019
Multimodal Model-Agnostic Meta-Learning (MMAML) code for learning few-shot classification. It contains
#PyTorch
implementations of our model and two baselines (MAML and Multi-MAML).
#NeurIPS
#code
I am delighted to share that our papers “Learning to Synthesize Programs as Interpretable and Generalizable Policies” () and “Generalizable Imitation Learning from Observation via Inferring Goal Proximity” are accepted by
#NeurIPS2021
!
The histogram of
#ICLR2020
paper ratings:
1. a paper with ratings 6, 8, 8, got rejected:
2. a paper with ratings 1, 3, 3 got accepted
3. 62% of 3, 6, 8 papers got accepted
4. only 27.8% of 3, 6, 6 papers got accepted
4/n While this might not be comparable to plagiarism, I thought this could raise people's awareness. Please do not use the code of the papers assigned to you for review for non-reviewing purposes, especially without permission from the authors.
For those who are anxiously waiting for the decision on ICML 2023 submissions, it probably won't come out soon. The decision notification time was secretly updated to Apr 24 '23 (AoE) on the ICML website. It was Apr 22, according to the archived webpage
I am proud to present my very first last-author paper as an Assistant Professor at NTU. I'm incredibly fortunate to work with these dedicated and talented students and colleagues: Guan-Ting Liu, En-Pei Hu, Pu-Jen Cheng,
@HungyiLee2
A thread 👉
This went beyond my expectation.
The first author of paper B (Duo Li) just posted on Zhihu (Chinese Quora).
Summary:
1. The literature survey was done before paper A posted on arXiv
2. The two papers are coincidentally similar since
Again, this went far beyond my expectation. So many plot twists.
Duo Li has completely revised his response to:
1. His ICCV 2021 paper and CVPR 2020 papers plagiarize other papers
2. His co-authors were not involved in plagiarism
3. He feels sorry and will learn from it
This went beyond my expectation.
The first author of paper B (Duo Li) just posted on Zhihu (Chinese Quora).
Summary:
1. The literature survey was done before paper A posted on arXiv
2. The two papers are coincidentally similar since
My very first NLP paper got accepted to EMNLP 2023! This is joint work with
@lunweiku
@windx0303
and the leading student author Nicholas Collin Suwono, which explores the Location-aware Visual Question Generation task, aiming to ask engaging questions from a specific location.
I posted my
#ICLR2020
rebuttal around two days ago and two of my reviewers have already replied with constructive comments. I’m extremely grateful. I replied to all the rebuttals of the papers that I am reviewing within 6 hours after the rebuttals were posted.
When will the
#ICLR
reviews come out? I am so anxious. I even wrote a script that checks my submission every minute and will send me an email once reviews come out.
We are extremely lucky to get reviewers who are responsible and kind - all of our reviewers have responded or updated the ratings.
The ratings of our two
#NeurIPS2021
submissions before/after rebuttal: 87776665/87777776.
I talked about "Learning to Synthesize Programs as Interpretable and Generalizable Reinforcement Learning Policies" hosted by
@hsuantienlin
at TAAI AI Forum 2023, covering our works that re-formulate solving reinforcement Learning tasks as program synthesis problems. 🧵
My Gmail app just found a hotel reservation in Hawaii from today’s arXiv update. Apparently,
1. Everyone is going to Hawaii for their holidays.
2. There's a hotel named Dong Yang Inn in Hawaii.
3. There's a paper author named Dong Yang from today's arXiv papers.
Thanks,
@GoogleAI
I will be attending ICML 2024
@icmlconf
in Vienna, Austria 🇦🇹! Happy to chat about diffusion models for imitation learning, programmatic RL, PhD in the US ➡️ faculty career in Taiwan, etc. Drop me a line or reply to this tweet if you are interested to connect. Please RT!
After 8 months of hard work as an Assistant Professor at National Taiwan University, a ton of effort in researching, teaching courses, mentoring students, and securing government and industry fundings, my lab finally can afford our first-ever robot, a robot lamp.
Since everyone on my Twitter timeline is attending
#NeurIPS2023
. I thought it would be helpful to share this Inigo Montoya conference networking tip for initiating conversations, which works great for me.
Po-Chen Ko
@PoChen23602
, an undergrad from our lab, who will apply for Fall 2025 PhD programs, will present our
#ICLR2024
@iclr_conf
work. If you are interested in learning from videos or are hiring PhD students, come talk to him at
#240
poster in the morning session on May 10.
This thread is outrageous. My advisor Prof Joseph Lim let me stay in Taiwan (even after I have been stuck here for two months waiting for my visa) and arranges all the lab meetings according to the time difference between Taiwan and Los Angeles.
Jung-Chun Liu
@JungChun1013
, a student from National Taiwan University and an incoming
@UMich
PhD student, will present our
#ICLR2024
@iclr_conf
work combining the strengths of deep RL and classical planning. Check out our poster in the afternoon session (4:30-6:30 PM) at
#270
!
Attending
@iclr2019
#ICLR2019
? Come to our poster at 11am on Wed: Composing Complex Skills by Learning Transition Polices. Our use previously acquired primitive skills to compose complex skills (obstacle course, tennis serve, etc.) by learning to smoothly transition between them.
After a long, anxious wait, the
#ICLR2020
result is finally out. We are so lucky to have all those hard-working reviewers reviewing our paper and the rebuttal.
#ICML2024
@icmlconf
Navigating through only ~15% posters in one poster session be like:
Going through all of them in 1.5 hours is really a cardio workout.
I will be presenting our work improving generative adversarial imitation learning (GAIL) by incorporating a diffusion model as a discriminator at the Generative Models for Decision Making workshop at
#ICLR2024
. Stop by our poster at 3 PM @ Lehar 3!
We are very grateful to our
#ICML
reviewers who rejected our paper but provided constructive feedback, which greatly contributes to our
#NeurIPS
paper. We sincerely hope that they see this message. Our ICML title was Model-Agnostic Meta-Learning for Multimodal Task Distributions.
The current cumulative sum plot of
#ICLR2020
submission scores is as follows. The average scores of 290 submissions have increased. The top 8 successful rebuttals are as follows (with an average score increase ranging from +4 to +2.33).
I reviewed three
#ICLR2020
submissions. For each paper, I spent roughly 6 hours reading the paper, 2 hours writing the review, and another hour reading the paper once again to make sure the points that I made in the review are valid. My reviews are 900-word long in average.
Many of the short reviews simply wrote something like “I don’t know this field at all.” I wonder why they couldn’t talk to ACs to get the paper re-assigned to other reviewers. Was it possible at all for this year ICLR?
Thrilled to share our
#COLM2024
@COLM_conf
paper "LLM Discussion: Enhancing the Creativity of Large Language Models via Discussion Framework and Role-Play"! This work explores enhancing LLM creativity via multi-LLM discussion and role-play!
A thread 👉
Our ICLR2019 paper Composing Complex Skills by Learning Transition Policies is released! We propose transition policies which enable agents to compose complex skills by smoothly connecting previously acquired primitive skills. See our demos and code at
I will be attending
@icml
! If you are interested in discussing research related to programmatic RL, program synthesis, robot learning, or just wanna meetup, my DM is open! I will be presenting two papers 🧵
For those who are traveling to Korea for
#ICCV2019
@ICCV19
: as far as I know, *Google map does not work in Korea* and the largest map provider (Naver) does not have an English version. Instead, you should download Kakao map:
Please stop telling undergraduates they need a first-author paper to get into grad school. It’s not true and puts way too much pressure and worry on them.
I always feel so sad when the papers that I am reviewing are withdrawn instantly after reviews come out. No discussion. No clarification. No answer to my questions. Nope.
My
#ICLR2020
preparation today:
- Got four vaccinations at once for yellow fever, hepatitis A, typhoid, and polio
- Got the medicine for malaria
- Registered at
Dear
#NeurIPS2021
reviewers,
Thank you for giving us a break since we posted the rebuttal. We are now mentally prepared for seeing your response.
Yours,
Authors
We will present our
#ICML
paper today at 2pm (
#116
), which re-formulates solving an RL task as composing a task-solving program that can be executed to interact with the environment and maximize the return.
If you are interested in RL or program synthesis, check out our poster!
3/n When I first saw that paper, I did not even know that crosses a line let alone doing something about it, but I definitely felt really uncomfortable. It has been a while and I still have not heard from the authors.