My student sent me this list saying they have to improve themselves in many areas. Such a list can do more harm than good. While I appreciate author's intention to motivate one for greatness, I don't think it can be planned. But you can plan to be a "good researcher."
Enjoyed visiting UC Berkeley’s Machine Learning Club yesterday, where I gave a talk on doing AI research. Slides:
In the past few years I’ve worked with and observed some extremely talented researchers, and these are the trends I’ve noticed:
1. When
Excited to share that I will be joining
@McGillU
and
@MILAMontreal
as an Assistant Professor in Computer Science and Linguistics from Jan 2020. Looking forward to work with excellent people who are passionate about human language and its role in learning! 🕺🏿🤸🏿♂️😊 1/3
Introducing WebLINX 🐯, a large benchmark for AI agents navigating real websites with multi-turn dialogue. 100K interactions across 2300 demonstrations on 150 real-word websites. Includes HTML, screenshots and videos. Tests unseen sites, tasks, blind users
Mindboggling that an entire paper of ours has been plagiarized and posted on arXiv. And
@arXiv
autodetect couldn't catch the plagiarism (and it took us a month). Is this a social experiment or a deliberate attempt? A nice example for the dark side of LLMs. Spotted by
@JasperJJian
Can we improve chatbots after deployment from natural user feedback? A big YES :) User feedback has rich cues about errors but it cannot be used directly for training. So we use GANs to generate training data from feedback.
#Findings
-of-
#EMNLP2020
#NLProc
.
I have an opening for a
#NLProc
postdoc at Mila/McGill.
Perks:
. pursue own interests
. mentor talented PhDs
. collaborate with Mila profs and industry
. access to large compute
. live in a vibrant city
Please spread the word. I will be in Toronto for
#ACL2023NLP
. Email me.
A glimpse of my journey into NLP and what I am excited about. Thank you CIFAR and
@FacebookAI
for your incredible support in funding fundamental research. Looking forward to hiring many junior researchers to work on hard problems in
#NLProc
at
@MILAMontreal
and
@McgillU
.
I will make my stance clear. I oppose any form of violence against civilians, especially children, be it in Israel, Gaza, Ukraine, Iran, Sudan or elsewhere. My silence does not mean I don't care. I care in limited ways in which I can make a difference while preserving my sanity.
In most search sessions, people switch between two to ten topics 😯. But work in conversational question answering assumes a session around a single topic (CoQA, QuAC). We launch, TopiOCQA (🗣️Tapioca🧋), a realistic challenge with topic switching
#NLProc
Mistral is not confused when we enable bidirectionality whereas LLaMA goes off the rails 🤠. We may have unlocked one secret ingredient of why Mistral is better than LLaMA. We believe it is 💥Prefix LM💥. This side finding is exciting in itself!
We also analyze how enabling bidirectional attention without training affects the representations of decoder-only LLMs 🔍. We find that Mistral-7B is surprisingly good at using bidirectional attention out-of-the-box 🤯 and speculate it was likely trained as a prefix-LM 🤔. 7/N
What are your tips for surviving grad school? We will share these with several grad students at
#acl2020nlp
.
My tip: commit to making at least one line of progress (code, writing) every work day even on your bad days.
#NLProc
#AcademicChatter
Transformer language models are good at reasoning eg if x is a parent of y, and y is a parent of z, they can infer x is a grandchild of z. But can they reason about long number of proof steps? No. See results when trained on 2,4,6 proof lengths and tested on 2-10
#NeurIPS
#NLProc
Back-training, the back-translation technique for non-translation tasks is here. We use it for unsupervised domain adaptation.
We see 9 BLEU points improvement on Question Generation, and 8 points on retrieval performance (P
@1
)
#NLPRoc
As a student, I find it paralyzing to aiming for greatness because you never know when you met the bar. Instead one should aim to master the craft of research -- a) ideation, experimentation, and communication, and b) graduate and get ready for a research career
School of Computer Science at McGill University
@McGillU
is hiring a tenure-track faculty position in
#MachineLearning
. The candidate may also be nominated for a generous CIFAR AI chair at
@Mila_Quebec
. Join us in Montreal!! Deadline Dec 15th. Details:
Using ~25 lines of code 💻 you can easily leverage retrieval-augmented instruction-following models for question-answering! 🚀
🔗
Detailed documentation coming soon! Shoutout to
@xhluca
,
@ncmeade
, and
@ParishadBehnam
for core contributions to the library
Today I will be talking at
#SPNLP
at 15:30 pm Dublin time on "Do we still need inductive biases after Transformer language models?". Come with your answers :), and questions are welcome too.
#acl2022nlp
LLMs are "secretly" powerful text encoders. LLM2Vec is the key to unlock their embeddings in 1-2 hours in an unsupervised fashion using LoRA. Achieves SOTA on MTEB in the unsupervised category and also among supervised models trained on public data
Code:
We introduce LLM2Vec, a simple approach to transform any decoder-only LLM into a text encoder. We achieve SOTA performance on MTEB in the unsupervised and supervised category (among the models trained only on publicly available data). 🧵1/N
Paper:
General sense I am getting out of my
#EMNLP2020
submissions:
1. Reviews are unnecessarily critical (an indication of first-time reviewers?)
2. Reviews are short
3. Many borderline scores
4. Low quality reviews (heavy reliance on paper matching?)
Pleased to have
@ChrisGPotts
's talk today at
@McGillU
/
@Mila_Quebec
on inducing causal structures to enforce systematicity. The key idea is to swap intermediate representations such that the prediction for the intervention conforms to the expected output. Works well for many tasks
Vision and language models have a long way to go. When context really matters,
@OpenAI
CLIP is only slightly better than random. ImageCoDe is probably the wildest dataset ever for pragmatics with natural images! Human performance is ~90% and the best model is at ~30%.
#acl2022nlp
Can vision & language models retrieve the correct image from a set given its contextual description (e.g. No bridesmaid visible at all)? We show that models struggle with this kind of contextual reasoning
#ACL2022
Debiasing methods may seem effective at mitigating bias but the trick they rely on is to make the language model worse. They do not generalize beyond gender. An empirical analysis of several debiasing methods on multiple domains and datasets, with a unified codebase!
#acl2022nlp
Social bias mitigation techniques in
#NLProc
require a holistic evaluation (bias, language modeling & downstream)!
Work w/ Elinor Poole-Dayan &
@sivareddyg
at
#acl2022nlp
.
📄Paper:
🔧Code:
📋Website:
1/7
At least an exception for people affected directly will be a kind gesture from
@EMNLPmeeting
! I am not affected personally, but I can tell you that the situation is dire for many. Over the past few days, most calls with my family end up with news of someone I know dying
#NLProc
Wrote to the EMNLP 2021 PC Chairs considering extending the deadline due to COVID-19 "catastrophic" situation in India, but was turned down. I recall in EMNLP 2020 the deadline was extended due to protests in the US, which caused some controversies. Submitting to CIKM 2021 now.
Question answering models don't have to die once deployed. Revive them with interactive feedback for an ever improving model both on accuracy and explainability. Introducing FeedbackQA
#acl2022nlp
Try the demo to see it yourself
#NLProc
Omar Khattab (
@lateinteraction
) is enlightening us at Mila/McGill with Language Model Programming, an important step towards controllable language models.
#NLProc
@stanfordnlp
Most common method for cross-lingual transfer is to use multilingual encoders instead of translations (translate to English and then do the target task). We find the latter is superior if you directly optimize the translations for the target task & use n-best translations
#NLProc
I might finally see some of my students physically. My heart breaks to see some of them attend a reading group at 3am. Grad life is hard, and this pandemic is making it harder. A reminder to be kind to our students ♥️
Starting in late October, international students attending a designated learning institution that has been identified by their provincial or territorial government as having a COVID‑19 readiness plan will also be able to enter Canada.
4. Disagree. Deep learning could not have happened if it was not pursued by the few people in their rabbit holes during the AI winter. This is not for everyone and could be highly stressful.
Super excited about your decision, thank you! Looking forward to doing important work in the space of language understanding and grounding. Some of the reasons you mention strongly resonate with why I started my research group at
@Mila_Quebec
.
Seems like an eternity since I decided to apply for a PhD farther away than I've ever been.
I can finally announce, very excitedly of course, that I'll join
@sivareddyg
's lab at
@Mila_Quebec
& collaborate with other language grounders!
Thread on things I'm excited about: 1/n
One of the qualities I appreciate of students is to having the ability to reach the finish line of a project. It is easy to hop on multiple projects but requires strength to stay. But I don't mean staying on for too long. Just the right amount. Find your balance :).
Along this process, it is fine to have papers that are not great but are scientifically rigor. As you gain more experience, you can aim to maximize impact but it is not guaranteed.
Finishing up
#EMNLP2023
after presenting some cool papers from our lab. It was nice catching up with friends and forming new connections. See you all at the next one!
How biased are large pretrained language models? Our take on measuring bias in existing LMs. We found that GPT2 is more humane than BERT and others. Terrific work by
@moinnadeem
(started his work as an undergrad)!!
Cautionary warning of offensive examples..
As pretrained language models grow more common in
#NLProc
, it is crucial to evaluate their societal biases. We launch a new task, evaluation metrics, and a large dataset to measure stereotypical biases in LMs:
Paper:
Site:
Thread👇
Many of us at
@Mila_Quebec
are thrilled to hear from
@HinrichSchuetze
about generating large scale instruction data in an unsupervised fashion. Recording will be available. My course students also had a bonus course lecture on pattern-exploiting training (PET) and GNNavi.
One of most intriguing findings of 2023 is that adversarial triggers that jailbreak one or more LLMs transfer to other models. We were so excited that we spent many months figuring out the conditions for universal transfer but the transfer never happened. It wasn't a bug 😀
Adversarial Triggers For LLMs Are 𝗡𝗢𝗧 𝗨𝗻𝗶𝘃𝗲𝗿𝘀𝗮𝗹!😲
It is believed that adversarial triggers that jailbreak a model transfer universally to other models. But we show triggers don't reliably transfer, especially to RLHF/DPO models.
Paper:
Really excited that
@DBahdanau
has joined
@McGillU
as an adjunct prof, a great strength to our NLP group. Through collaborations with my students, I know how much he cares about the growth of junior people working with him. He will be a fantastic supervisor, please apply
#NLProc
I am excited to share that as an Adjunct Prof at
@mcgillu
and member of
@Mila_Quebec
, I am looking to take 1-2 fully-funded MsC or PhD students this Fall. How to apply: (read carefully!). For possible research topics, see the thread.
We submitted a proposal to organize a workshop on document-grounded dialogue and conversational question answering. Excellent line of speakers/panelists and PC! We also have a shared task. More details
#NLProc
A systematic organization of posthoc interpretability methods in
#NLProc
. Result of many human hours (much more environmental friendly 🙂) from
@andreas_madsen
. It is likely that some relevant papers are missing. Please point them in order to be included in the next revision.
Our new survey on post-hoc interpretability methods for NLP is out! This covers 19 specific interpretability methods, cites more than 100 publications, and took 1 year to write. I'm very happy this is now public, do consider sharing.
Read . A thread 🧵 1/6
08 Feb
@StanfordNLP
seminar: Pararth Shah from Google on "Building a Conversational Agent Overnight with Dialogue Self-Play"
#NLProc
Please join us. Open to all!
Incredibly proud of Edoardo! Not only a brilliant person but a compassionate one. Will be an amazing supervisor, highly recommend! Lucky to have him in my group. This news sets a high bar for current members. Thanks to
@AnnalKorhonen
for sending one of her best students my way :)
I am delighted to share that I will be joining
@EdinburghNLP
at
@EdinburghUni
from 2022 as a lecturer in Natural Language Processing. I am currently recruiting PhD students, so if you are passionate... (1/6)
@strubell
@ACL2019_Italy
@ananya__g
@andrewmccallum
It is rare to see such papers and I admire your courage for choosing to work on such a problem. Congratulations! Every experiment I run from now on will make me rethink twice. Thank you!
5. Disagree. If you are a leader in an area, it is not wrong spending more time in that area while also pursuing other things if it is off the trend. Also, a degree like PhD requires you to come up with a coherent thesis, not papers on different topics stapled together.
2. Disagree. It is fine to have pieces of code that are hacky in the ideation phase. Later on when you get a validation of an idea, spending more time is justifiable. In industry, this could be different as there are many processes like code reviewing etc.
I am at
#EMNLP2023
to catch up with friends, peers and prospectives. DM me. My wonderful students are presenting:
#1
LLMs entangle many linguistic phenomena.
@JasperJJian
designs a clever way to extract syntactic structures from first principles.
How can we extract unsupervised syntactic structures from language models that are, well, syntactic?
I’ll be presenting 'Syntactic Substitutability as Unsupervised Dependency Syntax' (SSUD) at
#EMNLP2023
, on Dec 10 at 9AM in West 3! ✈️ This is work with
@sivareddyg
.
🧵
industry != academia. Once you find your ground, it is easier to take risks and aim for impact. You can still be a good researcher even if your papers are not ground breaking.
Coming to the list.
1. Agree. It is important to look at data before thinking about modeling. The first thing to try should be simple, and simple solutions are great solutions. It is also important to observe the behavior of the models qualitatively, and not just quantitatively.
On evaluating the faithfulness of explanations like attention and gradients -- the idea is to retrain a model by removing allegedly important tokens and measure the drop in performance, inspired from ROAR in vision. More faithful an explanation more the drop
#NLProc
Can you trust attention explanations in
#NLProc
? What about other explanations? I'm excited to present our answer to these questions in a new paper:
I'm proud of this work. I hope you find it useful. If you do, consider sharing or ❤️. What do we find?...
3. Disagree. Working independently or in a small group is perfectly fine. But research can be isolating, so leaning on others is a good idea. Make sure you master the craft of each step at least once in different projects. Too many collaborators could also be futile.
If you are looking for an NLP problem that is real and not solved by LLMs 🤖, check this out. Real conversations with Statistics Canada agents helping you to find relevant information from thousands of tables containing census data. It can't get more real than this 🙂.
#NLProc
Happy to share our
#eacl2023
paper, which introduces a new dataset for Table Retrieval in Conversations: the 🍁ꜱᴛᴀᴛᴄᴀɴ ᴅɪᴀʟᴏɢᴜᴇ ᴅᴀᴛᴀꜱᴇᴛ 🍁
Paper:
Homepage:
A work with
@sivareddyg
and
@harmdevries77
🧵👇
Excited to be the opening speaker at the AI Assistant Summit, SFO. I will be talking about the state of NLU, what's hype and what not, at 9:10 am today. Slides will be posted shortly.
#reworkdl
#reworkai
#nlproc
Job search has been a roller coaster of emotions. I am glad to have met many incredible people during the process. The toughest part was to bear the guilt of getting paid as a postdoc while I am busy searching and not working. I am grateful to have a very supportive advisor. 2/3
10 May
@StanfordNLP
seminar:
@Alan_Ritter
from Ohio State University on "Large Scale Learning for Information Extraction"
#NLProc
Please join us. Open to all!
12 April
@StanfordNLP
seminar:
@AdjiBoussoDieng
from Columbia University on "Deep Sequence Models: Context Representation, Regularization, and Application to Language"
#NLProc
Please join us. Open to all!
If you have any readings on how to get ready for an Assistant Professor job, please share. I have 3 months break to get started. What should I prioritize during this period --- travel, research, course preparation, catch up on tv, rest, or cook :). 3/3
If you are attending
#ICLR2019
, don't miss our workshop on Deep Generative Models for Highly Structured Data (NLP, Vision, Physical Sciences). An excellent line up of invited speakers (e.g., Yoshua Bengio), and awesome papers
#ICLRDeepGen
Tip 2: Meet your supervisor regularly (even for 5 mins) especially when you feel you are not progressing and don't have enough updates. Intrinsic motivation is hard to rely on. Meeting supervisor was my extrinsic motivation (or fear of embarassing myself which happens sometimes)
11 Jan
@StanfordNLP
seminar: Pushpak Bhattacharyya from
@iitbombay
on "Sarcasm Detection: A Computational Cognitive Approach"
#NLProc
Please join us. Open to all!
@sleepinyourhat
I sympathize with them partly. I just saw someone who almost had a panic attack realizing that they are the only one in the QA session. Icebreaking can be done from the authors' end. Good starter question is "Where are you from?, How do you find the conference?" and small talk.
08 March
@StanfordNLP
seminar: Michael Auli from Facebook AI Research on "Sequence to Sequence Learning: Fast Training and Inference with Gated Convolutions "
#NLProc
Please join us. Open to all!
I will make my stance clear. I oppose any form of violence against civilians, especially children, be it in Israel, Gaza, Ukraine, Iran, Sudan or elsewhere. My silence does not mean I don't care. I care in limited ways in which I can make a difference while preserving my sanity.
EMNLP/*ACLs assess the interest of a workshop through a survey (😐). If you work in the areas of dialogue/conversational question answering, please show your interest through votes here (scroll to the 5th page, category: discourse, dialogue).
#NLProc
We submitted a proposal to organize a workshop on document-grounded dialogue and conversational question answering. Excellent line of speakers/panelists and PC! We also have a shared task. More details
#NLProc
GOP tax plan will reduce annual stipend of PhD students by 25 percent, e.g., at CMU from $29,566 to $22,191. While Canada is investing millions in research, USA is taking away food from PhD students.
@nlpmattg
I listened to many voices on podcasts. But I did not anticipate a day where I would listen to my own voice. Thanks
@nlpmattg
and
@waleed_ammar
for making this possible! This is so much fun, and a bit strange :).
LLM2Vec meets Meta-Llama-3 → new SOTA among models trained on public data 🥇. We applied our LLM2Vec approach to Meta-Llama-3-8B and it works like a charm. Each step of the LLM2Vec pipeline improves the model's performance on embedding tasks 👇. 1/N
Models:
Belated but thrilled to welcome David to join the NLP faculty at McGill. He will be recruiting several students, a great opportunity to work on low-resource languages.
Thrilled to announce that I will be joining the
@mcgillu
School of Computer Science as a Tenure-Track Assistant Professor & Core Academic Member
@Mila_Quebec
starting early next year. A big thank you to my advisor, mentors & friends. God has surrounded me with amazing mentors.
Incredible effort by
@xhluca
and
@ZdenekKasner
in pulling off this behemoth task of building everything from scratch. They could have raised millions in funding for a startup, but instead, they decided to write a 40-page academic paper and release all data and code publicly.
@sivareddyg
#acl2020nlp
Tips for surviving grad school:
Chris Manning: achieving SOTA should not be the goal of Grad students. It's about choosing a niche area and come up with small useful ideas.
I successfully defended my PhD thesis yesterday 🎓. Many thanks to my advisor Dietrich Klakow, my committee
@JonathanBerant
, Vera Demberg, and Mareike Hartmann as well as my amazing colleagues and collaborators 😍. Next step 👉
@Mila_Quebec
with
@sivareddyg
🚀
.
@ReviewAcl
Is it true that in order to be a reviewer, one should already have 5+ published papers at NLP venues? My student with multiple *ACL/NeurIPS papers kindly volunteered but came to know they are ineligible. How is this scalable as most submissions are from students?
Dataset biases can easily overpower inductive biases giving us a false impression of progress. It is critical to have both in-distribution hard splits and out-of-distribution splits to measure generalization. We examine language grounding 1/n
Are SOTA models grounding the language correctly? Are the benchmarks evaluating them right? No! We observe that neural module networks (NMNs) are simply relying on input words instead of exploiting structure. ViLBERT is relatively robust (1/4)
#acl2020nlp
I spoke at the Bay Area NLP & ML meetup a few weeks ago. Today this arrived at my door. Such a kind gesture. Thank you
@MyaSystems
,
@sanjana_rampi
, Zac for making my day and organizing this event! Great reception & friendly audience. I recommend my
#NLProc
friends to speak here🙂
Thrilled to share that starting Jan 2022, I'll be an assistant professor at the University of Montreal and a core academic member of
@Mila_Quebec
!
I'm extremely grateful to my amazing collaborators and mentors, especially
@lgetoor
and
@blei_lab
!
(1/2)
#NLProc
folks: Is there a way to access past rebuttals on softconf (for our own submissions)? This is valuable data to train future students, and I am not sure why conferences hide it after the review period. At least I can see reviews which was not the case until recently (2021)
@sleepinyourhat
The expectation from authors and participants that they have to watch the video beforehand is unrealistic. I find it liberating to give up on that idea. Just go to a session and ask what the main contribution is or listen to other people questions.
Google Deepmind presents Transformers Can Achieve Length Generalization But Not Robustly
paper page:
Length generalization, defined as the ability to extrapolate from shorter training sequences to longer test ones, is a significant challenge for language
World affairs are depressing and I find it is important to preserve my mental sanity to make an meaningful impact. Otherwise, I find everything in my life goes downhill (including family time) and any efforts thereafter are a lost cause.
Let a bot be a bot. Let it acknowledge what it doesn't know but still be helpful and creative. Current dialogue datasets encourage the opposite and license models to hallucinate alarmingly (
#naacl
). FaithDial (preprint) is all about trustworthy chatbots