Giving a new talk tomorrow at
@USC
𝑳𝒂𝒏𝒈𝒖𝒂𝒈𝒆 𝑴𝒐𝒅𝒆𝒍 𝑨𝒍𝒊𝒈𝒏𝒎𝒆𝒏𝒕: 𝑻𝒉𝒆𝒐𝒓𝒚 & 𝑷𝒓𝒂𝒄𝒕𝒊𝒄𝒆
hosted by
@mahdisoltanol
While the talk is mostly an 𝑎𝑙𝑖𝑔𝑛𝑚𝑒𝑛𝑡 tutorial, I will also touch upon some of our recent work👇
The question that a reviewer should ask themselves is:
Does this paper take a gradient step in the right direction? Is the community better off with this paper published? If the answer is yes, then the recommendation should be to accept.
[
#eacl2024
paper]
TL;DR We introduce 𝗴𝗿𝗮𝗱𝗶𝗲𝗻𝘁-𝗯𝗮𝘀𝗲𝗱 𝗿𝗲𝗱 𝘁𝗲𝗮𝗺𝗶𝗻𝗴 (𝗚𝗕𝗥𝗧), an effective method for triggering language models to produce unsafe responses, even when the LM is finetuned to be safe through 𝑎𝑙𝑖𝑔𝑛𝑚𝑒𝑛𝑡.
Please RT
Are you a *PhD* student conducting research on generative models? Are you excited about
#ResponsibleAI
aspects of generation? We are looking to host a student researcher/
#internship
in 2023. Please get in touch via *email*.
P.S. I'll be at
#NeurIPS2022
and
#emnlp2022
Have you been perplexed by the surprising performance of 𝗯𝗲𝘀𝘁-𝗼𝗳-𝗻 in alignment compared to SOTA method (𝗣𝗣𝗢/𝗗𝗣𝗢/𝗜𝗣𝗢)?
We have theory that explains this phenomenon.
Both
@icmlconf
and
@NeurIPSConf
held in the US in 2022-23! The US is one of the most visa unfriendly states (appointment wait times 6+ months, processing another 6+ months), this is significantly hurting diversity & inclusion. We should strive to do better!
#ICML2023
#NeurIPS2023
We publish a ton of our core work on language models. Here are a few samples just from our team released in the last few weeks:
-
-
-
-
How's that "giving almost nothing back"?
I can't help being a bit sad in thinking how much the Gemini release is taking from the academic and open LMs communities, while giving almost nothing back
I can't help but wish our relationship with commercial players would be more of a two way channel
Have you been compiling reward-KL tradeoffs to compare different alignment methods?
Have you been using 𝐛𝐞𝐬𝐭-𝐨𝐟-𝐧 as a baseline?
Have you wondered about the analytical formula that claims this formula?
𝐾𝐿 (𝑏𝑒𝑠𝑡-𝑜𝑓-𝑛 || 𝑏𝑎𝑠𝑒) = 𝑙𝑜𝑔(𝑛) - (𝑛-1)/𝑛
I left Meta AI a few weeks ago. I am filled with gratitude for the opportunity to “think big” and propose Project CAIRaoke to redefine the future of
#ConversationalAI
! I am also humbled to have worked with so many amazing people along the way!
[PSA]
Please use \𝗰𝗶𝘁𝗲𝘁{} and \𝗰𝗶𝘁𝗲𝗽{} correctly or the paper will be hard to read.
\𝗰𝗶𝘁𝗲𝘁{}: Author name is intended to be part of the sentence: X et al. (2024)
\𝗰𝗶𝘁𝗲𝗽{}: Citation appears in parentheses and not read in the sentence: (X et al., 2024)
Off policy algorithms (even the simplest ones) are indeed useful for RLHF and can train a value function on massive amounts of data and 𝗼𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺 𝗜𝗣𝗢/𝗗𝗣𝗢/𝗣𝗣𝗢.
A brief summary on what REINFORCE is in terms of RLHF and history of RL.
The algorithm known as REINFORCE is really just the vanilla policy gradient approach. The name comes from Williams 1992, "Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement
Excited to share that our paper titled “Winning is not everything: enhancing game development with intelligent agents,”
@IEEETxnOnGames
, June 2020, has been selected to receive the 2023 Outstanding Paper Award by
@ieeecis
Awards Committee.
A periodic reminder to reviewers:
If you ask authors for more experiments, then you need to communicate a clear hypothesis you're trying to verify with those (e.g., effectiveness on imbalanced data, generalization beyond a certain modality, scalability, etc).
Otherwise don't!
As for next steps, I am excited to share that I have joined
#GoogleResearch
to lead efforts around robust and fair development of core machine learning techniques. I am also moving to New York City, and excited to be back to the east coast!
If you reviewed for
#NeurIPS2023
, please take a moment to read the authors' rebuttal, other reviews, and briefly respond to the authors (even if the response is that I need more time to process your rebuttal). The authors put a lot of time into this!
I've seen variants of this terrible idea proposed before.
Not only it's plain wrong and unethical to not provide recommendation letters to top students, it also doesn't work! Students will find other ways to leave, and then they will not even look back!
When I first started reviewing ML papers, I was fighting to reject bad papers. These days I find myself fighting to accept good papers.
The change in perspective has also made my reviews more constructive even in cases where the recommendation has to be reject.
An O(n) algorithm (including random sketching) to check out the posters is no longer feasible!
Does anyone know a sublinear algorithm for retrieval/learning from an unsorted knowledge base?
#NeurIPS2023
Here is what I "speculate" might have happened at
#emnlp2023
:
After decisions finalized by AC/SACs, the
@emnlpmeeting
PCs likely felt there were "too many" accepted papers (in findings), and scrambled to reject 100s of papers to lower the acceptance rate to keep "prestige".
The Conversational AI Research team at
@facebookai
is seeking to recruit multiple PhD student interns who are doing research in the conversational AI domain! Let's meet and chat tomorrow at the Conversational AI workshop!
#neurips2019
It was an honor to receive the
@IEEETxnOnGames
Outstanding Paper Award for our 2020 paper titled "Winning Is Not Everything" at
@ieee_cog
2023, from
@togelius
tonight!
If you are at
#EACL2024
, on Friday 4pm Malta time, I will give a (virtual) talk on 𝗖𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗲𝗱 𝗗𝗲𝗰𝗼𝗱𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 at the 1𝘴𝘵 𝘗𝘦𝘳𝘴𝘰𝘯𝘢𝘭𝘪𝘻𝘢𝘵𝘪𝘰𝘯 𝘰𝘧 𝘎𝘦𝘯𝘦𝘳𝘢𝘵𝘪𝘷𝘦 𝘈𝘐 (𝘗𝘌𝘙𝘚𝘖𝘕𝘈𝘓𝘐𝘡𝘌) 𝘸𝘰𝘳𝘬𝘴𝘩𝘰𝘱.
Dear authors: It is NOT okay to bully the reviewers who have engaged with you in a discussion and don't agree that your paper should be published! Thanks!
#NeurIPS2021
If you are a *PhD student* interested in Conversational AI Research, and would like to do an internship at
@facebookai
in Fall 2020, please reach out via email!
#internship
#ConversationalAI
There are two types of empirical researchers:
(1) Devise a marketable narrative. Cook up experiments to support the narrative. Write a paper!
(2) Devise a hypothesis. Design experiments to verify/disprove the hypothesis. Write the narrative based on results or drop the project.
Headed to Vancouver to
#AAAI2024
, giving 2 talks & serving on a panel! Hope to see you there!
𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁: 𝗧𝗵𝗲𝗼𝗿𝘆 & 𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗲
While the talk is mostly an 𝑎𝑙𝑖𝑔𝑛𝑚𝑒𝑛𝑡 tutorial, I will also touch upon some of our recent work👇
It's completely acceptable to write a purely empirical paper as long as the empirical results substantiate the claims!
It's not ok to inject nonsensical/wrong math "results" into an empirical paper to give the illusion of having supporting theory! This happens way too often!
[Call for papers]
#NeurIPS2023
R0-FoMo Workshop
Robustness of Zero/Few-shot Learning in Foundation Models We solicit novel contributions that relate broadly to zero/few-shot learning in foundation models, with both empirical and theoretical nature.
Please R/T
Identify and remove the bad ACs! There is no shortage of qualified people who can AC!
Let the authors and reviewers identify and report these bad ACs, and don’t invite them back!
Going to San Diego to attend the ITA workshop (), which is my favorite conference/workshop of every year!
I will speak about our recent findings on 𝐥𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐦𝐨𝐝𝐞𝐥 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 on Friday!
Are you excited about your new 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 method with a higher win-rate against the base model? Are you ready to publish?
WAIT!
Unless your judge was perfect (which defeats the purpose), comparing win-rate without regard to the divergence from the base policy is flawed!
I will be at
#NeurIPS2023
and look forward to meeting old and new friends!
Would love to connect and chat about building safe, helpful, and scalable generative language models!
If you reviewed for
#ICML2024
, please take a moment to reflect on the author rebuttal, other reviews, and ask any further clarifying questions that authors can help answer.
The authors' ability to engage in discussion ends in less than 24hrs.
And the
#icml2023
test-of-time award goes to "learning fair representations" which has arguably had a huge impact on ML. Congratulations to the authors!!
Check out our
#EMNLP2023
paper on evaluating and improving diversity of representation in LLMs!
P.S. We're looking to host interns to work on different RAI aspects of using LLMs. Stay tuned on more details!
Large language models (LLMs) have come a long way and solve many tasks. BUT diversity and inclusion in LLM generations is still an open challenge.
📢 New
@emnlpmeeting
paper on quantifying and improving people/culture diversity in LLMs:
#EMNLP2023
1/n
Check out Somnath's
#ICLR2024
spotlight paper on how to improve fairness in online learning.
TL;DR Using an oblique decision tree and imposing fairness constraints on each node helps achieve good fairness guarantees empirically/theoretically!
See you in Vienna to discuss more!
How can we improve fairness when data arrives in an online stream?
We present 𝐴𝑟𝑎𝑛𝑦𝑎𝑛𝑖, an effective approach towards achieving group fairness in online environments.
𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭 paper at
#ICLR2024
. Work from my last internship at
@GoogleAI
.
Happy Women in Mathematics day everyone!
May 12 is the birthday of Maryam Mirzakhani, a mathematician who was awarded the Fields medal ( highest honor in math) for her contributions to geometry and dynamical systems.
[Two of my fav mathematicians: Maryam and Ingrid Daubechies]
Please distinguish the citizens of a country from its government. Targeting the citizens of a country for the aggression of its government is not the solution!
@JonathanBerant
I'm really sorry for how you're feeling, Jonathan!
I thought regardless of how one might think of Israel's government, it should be a very low bar to condemn Hamas' terrorist attack against civilians. I still think/hope that's the case for majority of the community.
Robustness methods
1) augment data with natural/synthetic perturbations and a consistency loss
2) reweight samples to improve generalization (like DRO)
We do it differently!
We show significant robustness with a simple tweak of the first layer and loss motivated by comms theory.
📢📢📢 Late post, but here we go...! I am thrilled to announce that our work on 𝙚𝒏𝙝𝒂𝙣𝒄𝙞𝒏𝙜 𝙤𝒖𝙩-𝙤𝒇-𝒅𝙞𝒔𝙩𝒓𝙞𝒃𝙪𝒕𝙞𝒐𝙣 𝙧𝒐𝙗𝒖𝙨𝒕𝙣𝒆𝙨𝒔 of deep neural networks has been accepted to 𝘼𝑰𝙎𝑻𝘼𝑻𝙎 2024!
I will be at
@aclmeeting
next week, and look forward to connecting with old and new friends!
We also have a student researcher position open in the fall (and potentially beyond) to study responsibility aspects (e.g., robustness, safety, fairness) of generative (language) models.
The question that a reviewer should ask themselves is:
Does this paper take a gradient step in the right direction? Is the community better off with this paper published? If the answer is yes, then the recommendation should be to accept.
Wondering what's going on in Iran? A glimpse of past 24hrs:
- A protester shot in the head for dancing w/o hijab
- A 3-year old detained
- Water shut to Kurdish city of Sanandaj to deter protests
World must stop engaging with this murderous regime NOW!
#مهسا_امینی
#Mahsa_Amini
I will be at
#NeurIPS2023
and look forward to meeting old and new friends!
Would love to connect and chat about building safe, helpful, and scalable generative language models!
[thread with
@meisamrr
and
@pooyaan
]
The
@NeurIPSConf
Experiment
We do a simple analysis that reveals that “if the
#NeurIPS2020
review process is redone with new AC/Reviewer assignments, the majority of the papers that are accepted today will get rejected.” 1/7
As
#NAACL2022
came to an end, the two most interesting personal takeaways:
1. I met for the first time so many people I had been working with over the past 2 years!
2. I learnt there are so many disciplines in
#NLProc
that I embarrassingly didn’t even know the problem statement
Check out Zhaofeng's work from his internship with us!
TL;DR A reward model trained on language S preference data could be used to align a language T LLM. This sometimes works even better than using a reward model trained on language T preference data.
Want to train an aligned LM in a new language 🌏 but don’t have preference data for training the reward model (RM)?
💡 Just use a RM for another language: it often works well, sometimes even BETTER than if you had a RM in your target language! 🤯
TL;DR We speed up speculative decoding (for free) using optimal verification algorithm for acceptance/rejection at block level!
This is complementary to numerous works that propose better drafting algorithms as all of them use the same token-wise verification (Leviathan et al).
𝗦𝗽𝗲𝗰𝘂𝗹𝗮𝘁𝗶𝘃𝗲 𝗱𝗲𝗰𝗼𝗱𝗶𝗻𝗴 is popular for speeding up LLM inference. Many have proposed better drafters, BUT they 𝑎𝑙𝑙 use a token-by-token verification procedure (𝘛𝘰𝘬𝘦𝘯𝘝𝘦𝘳𝘪𝘧𝘺).
𝘐𝘴 𝘛𝘰𝘬𝘦𝘯𝘝𝘦𝘳𝘪𝘧𝘺 𝘰𝘱𝘵𝘪𝘮𝘢𝘭?
Somewhat surprisingly, 𝗻𝗼.
I'll be at
#AISTATS2024
later this week!
With Madhow, we will co-present
@BhagyashreePu13
's poster on TEXP to improve robustness with a tweak to the first layer of the network.
Looking forward to meeting old and new friends!
📢📢📢 Late post, but here we go...! I am thrilled to announce that our work on 𝙚𝒏𝙝𝒂𝙣𝒄𝙞𝒏𝙜 𝙤𝒖𝙩-𝙤𝒇-𝒅𝙞𝒔𝙩𝒓𝙞𝒃𝙪𝒕𝙞𝒐𝙣 𝙧𝒐𝙗𝒖𝙨𝒕𝙣𝒆𝙨𝒔 of deep neural networks has been accepted to 𝘼𝑰𝙎𝑻𝘼𝑻𝙎 2024!
I have been involved in the review process of 4
@TmlrOrg
papers, and guess how many got in?
.
.
.
.
All 4 have been deservedly accepted, and it feels good!
Everyone pointed this breaching the professional agreement as a reviewer.
That aside, there are actual people behind the paper who put in actual work. Reviewer's job is to help them improve the paper and determine whether it's publishable. It's *never* to ridicule the authors!!
A 10-year-old girl is raped. The State forces her to remain pregnant and tells her to consider it an “opportunity.”
This isn’t Iran. This isn’t Gilead. This isn’t hypothetical.
This happened today in Ohio.
Is there a rigorous training for Reviewers/ACs/SACs in our community? Apparently not! This Metareview is completely unprofessional and unacceptable! Why didn't the SAC flag it?
I was searching through the opted-in NeurIPS rejections and found this rather rude metareview. The prior sentence is also a bit harsh, but the last sentence is totally unnecessary.
Please try to be more considerate than this when reviewing.
After reviewing for several
#ML
confs I had a personal reflection. I realized my bar for accept is much higher than accept rate. That led me to think harder to articulate reasons to accept, despite shortcomings! A paper need not be perfect, it should be a step in right direction.
One third of the rejected
#NeurIPS2023
papers in my stack would likely have had a chance with some additional results, revisions, reframing, and discussion, to solidify their pitch.
That is where
@TmlrOrg
wins!
I am excited to be attending
#ICML2023
this week, and look forward to catching up with friends and making new ones. I would also love to discuss responsible/reliable/efficient foundational models.
Please email me if you would like to meet 1:1
A 🧵on events I am attending Wed-Sat
Here is an idea!
If you are working on
#ResponsibleAI
and/or LLMs and have a paper in
#NeurIPS2023
poster sessions, please reply with your poster number and time slot so that I (and others interested in similar topics) don't miss it. Thank you!
An O(n) algorithm (including random sketching) to check out the posters is no longer feasible!
Does anyone know a sublinear algorithm for retrieval/learning from an unsorted knowledge base?
#NeurIPS2023
'On Tilted Losses in Machine Learning: Theory and Applications', by Tian Li, Ahmad Beirami, Maziar Sanjabi, Virginia Smith.
#tilting
#minimization
#risk
As a
#NeurIPS2023
author, you effectively have unlimited space to rebut the reviewers' response (by posting multiple 6k char comments).
Do NOT do that!!
If you want to win your argument, please strive to prepare a point-by-point response that is clear, convincing, and concise!
Iranian people are at war with a regime that has a track record of disregard for human life. It’s heartbreaking to see new lives taken in the streets of Iran by this murderous regime. We remain hopeful for the people’s will to prevail!! Be their voice!
#MahsaaAmini
#مهسا_امینی
Congratulations to all top reviewers of
#neurips2023
!
The quality of the program hinges upon the volunteer work of individuals who go above and beyond to fairly assess the claims and quality of the submitted papers!
The LM reasoning literature (e.g. CoT) is predominately focused on mathematical tasks with well defined responses.
What if we want to reason through more nuanced and complex topics with no clear right or wrong answer?
***Thought experiments promoting!***
Introducing 🤔 Thought Experiments prompting!
⛓️Chain-of-Thought is great but real world is messier than that, especially when answering challenging moral questions.
We show self-posing and answering counterfactual questions allow LLM to improve. 1/n
Happy Women in Mathematics day everyone!
May 12 is the birthday of Maryam Mirzakhani, a mathematician who was awarded the Fields medal ( highest honor in math) for her contributions to geometry and dynamical systems.
[Two of my fav mathematicians: Maryam and Ingrid Daubechies]
So many criticisms of
#NeurIPS
reviewer not knowing X or Y. This is mostly on the AC for not making sure that the assigned reviewers have the expertise to evaluate the work, and the SAC for not making sure that the AC has the expertise to identify such reviewers!
FERMI is the first stochastic method for group fairness with guaranteed convergence, with batch size as small as one sample!
More on the theoretical/empirical results, the paper, and the code package in the thread by
@BaharloueiSina
📢 🚨 Publication Alert🚨📢
In our recently published paper at
@TmlrOrg
with my fantastic collaborators Andrew Lowy, Rakesh Pavan,
@meisamrr
, and
@abeirami
, we propose the first provably "convergent" stochastic algorithm to achieve group fairness in large-scale machine learning.
[Unpopular opinion]
1. Do I think *CL anonymity should be abolished?
Absolutely. I think it hurts the very people it claims to help; any data showing otherwise?
2. Do I think the paper should've been desk rejected?
Absolutely. A rule is a rule and should be enforced uniformly
Just got a desk reject, post-rebuttals, for a paper being submitted to arxiv <30 min late for the anonymity deadline. I talk about how the ACL embargo policy hurts junior researchers and makes ACL venues less desirable for NLP work. I don’t talk about the pointless NOISE it adds.
Fun game. Clocking 17973 citations: "Distilling the knowledge in a neural network"
@geoffreyhinton
,
@OriolVinyalsML
,
@JeffDean
Reviewer 38 (NeurIPS 2014): "This work is incremental and unlikely to have much impact even though it may be technically correct and well executed."
🚨New paper!🚨
Self-Rewarding LMs
- LM itself provides its own rewards on own generations via LLM-as-a-Judge during Iterative DPO
- Reward modeling ability improves during training rather than staying fixed
...opens the door to superhuman feedback?
🧵(1/5)
I love when ML researchers get excited about science, but seriously the reviewing process for scientific applications at ML conferences (e.g., ICLR) is entirely broken.
Papers with glaring errors are sailing through, without a single review from somebody with domain expertise.
Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking
paper page:
Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language
As
#acl2023
came to an end, my main takeaway is that this community is extremely resourceful and very welcoming!! I've made many new friends and I'm glad that I've chosen this community as my second home (after information theory) to learn and grow!
7 papers in 21 days (15 working days)
-> 1 review every 2 days (16 working hours)
-> 1.5-2.5hr per review (assuming spending 10-15% of one’s time on
#NeurIPS2021
reviews).
Why do we expect people to spend more time on our
@NeurIPSConf
submissions?
If you are at
@icmlconf
, please come attend the ITR3 workshop on Saturday July 24!
The goal of the workshop is to explore information-theoretic methods for rigorous, responsible, and reliable machine learning.
I will be at
#NAACL2022
Monday-Friday! This is my first in-person
#NLProc
conference, and first major conference in more than 2 years! Looking forward to meeting y'all!
In industry, we mostly care about the candidate's vision. They need their published work to support the artifacts they describe in the vision.
Many papers that don't add up to a unified vision is certainly not a good sign.
The gap between perceived and actual incentives is pretty large. Number of papers or citations weigh pretty lightly in hiring discussions, in my experience.
(What does count are letters, which ideally say "Candidate solved important problem x, which is impressive because y")
🎉 Excited to announce that our paper has been ACCEPTED as a Spotlight Paper
@NeurIPSConf
!
Interested to know how video self-supervised learning (VSSL) methods behave under different types of real-world distribution shifts (e.g., in OOD)?
w/
@abeirami
@Ali_Etemad1
#NeurIPS2023