Work with
@nikhilchandak29
,
@DominikPeters
just won the Outstanding Paper Award🏆at AAAI 2024 (top-3/12000+). We study a framework for making multiple decisions that fairly satisfy diverse preferences. Potential applications range from faculty hiring to how AI combines values🧵👇
Students at
@iiit_hyderabad
, supposed to be one of the top engineering colleges in the country, are in an ongoing health crisis caused due to appalling mismanagement and negligence going on since well over a year. 🧵👇on mass typhoid breakouts, food poisoning, underreporting..
Students are forced to subscribe to the college 'mess' (apt word). Cockroaches in food, flies, lack of handwash etc. are just meant to be ignored, since years. The fact that there's less food than oil is somehow not even a major concern. Student complaints are ignored.
When students report sickness, hostel and health authorities blame it on food orders from
@swiggy
@zomato
, which frankly are much safer. It's so useless that students have given up and stopped even trying to report. Claiming plausible deniability is unfortunately a trend.
Luckily, this cover blew apart recently when
@OlympiadPanini
high-school students got severely sick and hospitalised, and they weren't ordering food. The worst part is, the college hid this from the rest of the student body, leading to dozens of avoidable sick students.
Being sick every month is just a part of Life
@IIIT
. For an institute that's proud of its CS research output, maybe it's time to look inward at the living conditions of the students driving this research. Won't even get started on codified exploitation in research practices.
I'm writing this here only because current students can't due to repercussions, and alumni are just relieved their time here is over. Maybe this stops faculty from turning a blind eye. My hunch is this is common in many top engineering colleges in 🇮🇳, but it really shouldn't be.
This isn't the first time. As
@pingiiit
reported, last year there was a widespread Typhoid breakout with 40+ cases due to contaminated water. The boys hostel warden intimidated students from getting tested, actively spreading false information about symptoms, worsening things.
This was met with coverups, scientific falsehoods and false promises, which often got exposed on internal mailing lists. New water coolers were installed, mostly in academic and research buildings, not changing the OBH 3rd floor cooler which caused most cases.
When
@sinha_shiven
told me he wants to work on AI solving IMO problems with Indian univ compute I was skeptical... Just a month later, so glad to be surprised! They almost matched AlphaGeometry with 0 GPUs, within 5 minutes. Questions the hype around LLMs for solving math
Excited to announce our preprint!
We develop a symbolic system for IMO Geometry that can rival Silver Medalists. Combined with AlphaGeometry, it outperforms IMO Gold Medalists in Geometry for the first time 🏅
Will be attending
@iclr_conf
and the
@DMLRWorkshop
in Vienna. Would love to chat about all things Alignment, Interpretability, Data-Centric AI, and how models (should) deal with conflicting training data, or any application that excites you :) DM/Reply
#ICLR
#ICLR2024
We extend our appreciation to the exceptional reviewers whose expertise enabled a high-quality selection process for
#DMLRWorkshop
@iclr_conf
2024!
Magdalena Proszewska
Yifan Zhang
Miguel de Benito Delgado
Agam Goyal
Shashwat Goel
David Esiobu
Ben Feuer
Ian Beaver
I’ve spent some time studying Russian and Chinese school programming culture.
The obvious: Do what China and Russia do - incentivise college admissions based on results.
Not so obvious: Create a better supply of teachers who can guide young students and scout top talent early on.
We demonstrate the use of unlearning to remove potentially harmful dual-use knowledge and capabilities from LLMs.
Was super cool seeing this play out from a small exploratory project during my time at
@MATSprogram
to such a large scale collaboration! See the TIME article link👇
The White House Executive Order on AI highlights the risks of LLMs empowering malicious actors in developing biological, cyber, and chemical weapons. To measure and reduce these risks, we’re releasing the Weapons of Mass Destruction Proxy (WMDP) benchmark.
(🧵below)
Surprising climate change is a major political issue everywhere in the world, but not one in India, despite the national capital facing some of its worst effects.
Asked GPT-4o to generate MCQ questions with answers from a blog for a quiz I'm making. 50%+ answers are c), and 80%+ are either b) or c). The ratio of b/c increases as I ask it to make more challenging questions. Possibly reveals something about the internet distribution of mcqs?
The idea of "machine unlearning" is getting attention lately. Been thinking a lot about it recently and decided to write a long post: 📰
Unlearning is no longer just about privacy and right-to-be-forgotten since foundation models. I hope to give a gentle
Things I never expected:
@AmyPrb
becoming hot in the pro-GOFAI community. Follow him for a very cool GOFAI result coming out soon. Spoiler: GOFAI matches AlphaGeometry with only a few minutes of CPU time
Oh no! LLaMAI under attack.. 😱
"multimodal models require exponentially more data to achieve linear improvements in downstream “zero-shot” performance"
So what if it is "exponentially more data"? We know offline data or compute complexity doesn't matter 🙄.. c.f.
🧑🏽🏫 👨🏽🏫Course project (list 👇🏽) poster presentation. CS7.405: Responsible & Safe AI Systems
@iiit_hyderabad
. Please join if this is of interest to you. Open to public / outside campus also.
Course materials:
#ProfGiri
#RAISpring2024
Thrilled to present my latest work on evaluating approaches for Machine
#Unlearning
-- data removal from trained ML models. Was exciting to leverage insights from Empirical DL Theory and Attacks on ML. Glad to be working with my amazing co-authors
@ponguru
, Ameya Prabhu.
@furongh
Our paper: may be of interest, we discuss how different preferences can be aggregated for proportional representation to different groups. This ensures decisions don't overweight contrarian individuals, something maximin can suffer from.
The same set of voters may keep on being satisfied. Some voters might end up approving no decision💔. Instead, if 30% voters agree (approve a common alternative) in 10 rounds, we want them to approve 3 or more decisions. Ideally, proportional influence even on worst-case inputs.
Suppose you wish to decide where to hangout with your group of friends. In each 'round', every 'voter' approves (👍/👎) some 'alternatives', and one alternative is to be picked as a 'decision'. Simple approach? In each round, pick the alternative with most approvals. Whats bad?🛑
PSA: Stop pretraining your VLMs on EN-filtered data, even if it improves ImageNet and COCO‼️
Doing so impairs the model's understanding of non-English cultures❗️
I argued for years, now finally publish concrete results for this (imo) intuitively obvious recommendation
A🧾🧶
Unfortunately
@nikhilchandak29
's Canada visa was not provided on time. Thankfully, Nicholas Teh () from Oxford graciously agreed to deliver the talk on Friday 2PM at Room 211 at AAAI on our behalf🙏. Drop by to learn more, and feel free to reach out to us!
For more details about our work, check out our video and paper. Work started during an internship at LAMSADE, Université Paris Dauphine-PSL. Special thanks to Jérôme Lang for inviting us and advising the project!
📹 📃
Pleased to be recognized as an Outstanding Reviewer (Top 10%) by
@icmlconf
in my first attempt at reviewing! Grateful for getting the opportunity as an undergrad considering I hadn't published at an ML venue before.
#ICML2022
We show an interesting implication for AI: Typical setup (eg: RLHF) of combining data from different groups and maximizing accuracy can lead to unfair outcomes, even for balanced datasets❌ Learning separate models for each group 🇮🇳🇫🇷🇺🇸 and aggregating with our rules does better!
@_akhaliq
explores drawbacks of maximizing utility from approval votes (it is highly majoritarian and ignores smaller groups), and presents ways to aggregate preferences with worst case guarantees on representation.
Our work opens many interesting directions for future work. Theoretically, can we extend from approvals to cardinal utilities? Applications: Democratic processes🧑⚖️; AI pursuing a mixture of (sometimes conflicting) goals 🤖; Aligning with subjective values 👨👩👧👧🇺🇳; and many more!
AI Safety researcher: "Let me come up with the most intricate pathways for AGI catastrophe to show how effective my research is"
AGI: "Sounds like a plan *rubs hands*",
The model is probably learning to he unsafe from AI Safety posts in the training data. Ironic.
Finally, I had to try out the paperclip test, since it's practically the Hello World of alignment at this point. Nice to know there will be a few humans left over!
@gaur_manu
Saw this, but it focuses more on answering MCQ questions rather than generation. Would be interesting to see if question generation is biased towards different options from question answering
We believe our work highlights massive scope for new research contributions, including new evaluations🧪, unlearning methods 🔨, and theoretical analysis 📝of the Corrective Unlearning setting. For more details: 📒 👨💻
@savvyRL
And yes, one of these was at ICML 2023, before I had a top-tier submission. It was a great learning experience, and Ig I didn't do a bad job? Or maybe the recognitions are really noisy too.
We study popular voting rules in 3 settings: online, semi-online (only number of rounds known apriori) and offline. We provide comprehensive results on which properties each rule satisfies✅, and reduce to a well known open problem❓in most remaining cases.
I've worked on academic deep learning and summarization for years.
Summarization is a foundational technology for the information age and a remedy for the attention economy.
Here's a🧵 for how we think and apply summarization at
@YouSearchEngine
@savvyRL
DMed. Tysm for doing this! Since there's backlash just thought I'd add my pov. I love reviewing. In the few times I got a chance, I've been recognised as an outstanding reviewer whenever there was one. I hope gatekeeping is not the solution for some of the very valid concerns :(
@kenziyuliu
Wanted to write something like this, and you've done an excellent version of what I had in mind! Really useful resource for people starting out in Unlearning, will share widely :)
@boknilev
I agree that probing is similar to representation reading and more citations to this work should be added. However, the unique takeaway is the extent to which using methods like PCA/mean difference (Figure 12) finds activation directions that can control generation/removal.
Specifically for removing the BadNet poison, only one method (SSD) studied succeeds, showing the tractability of generalizing removal from a representative subset. However, SSD hurts model utility, leading to significant drops in test-accuracy on clean samples.
We formalize this by adapting Justified Representation (PJR, EJR) axioms from Social Choice Theory literature. We prove our axioms are tight: solutions that satisfy stronger guarantees may not exist, and even when they exist cannot always be found online😓.