Dylan Sam @dylanjsam Twitter profile

Last Seen Profiles

@joker75z

@MarioleHanafi

@sabr_salah_1357

@wilemang

@Sam_Tomss

@NicoletteGuyana

@diozziro

@KhalilZaccaria

@stw_pdg

@INSquashCircuit

@bokeplokalmalam

@Recreinm

@Sportsdesignind

@mikurikooo

@Recreinm

@atienospirit

@TLAB_ke

@UltimateTroop

@stw_pdg

@FELlnwza

@NCT_INDONESIA62

@BikesNotCarsLA

@Tromcx

@pekuyocuhcuh

@OTO_iwateh

@bwmr652143

@gatlandsquad

@BjornNordnet

@shawnpics

@JoeJoyce94

@JeetSinghIN

@voxelqueen

@QuirkyLogic

@bokeplokalmalam

@Indo_Nakall

@iilnue

Dylan Sam

@dylanjsam

4 months

I started a summer internship at @GoogleAI in NYC this week, where I’ll be working on data for LLMs! I’d love to meet up with friends or researchers in the area 🙂

10

197

Dylan Sam

@dylanjsam

3 years

I'm super excited to share that I'll be joining @mldcmu and @SCSatCMU as a ML PhD student this Fall! It's been a crazy application season, and I'm looking forward to diving into more ML research!

8

1

106

Dylan Sam

@dylanjsam

3 years

First week of classes done @CarnegieMellon ! Moved into my new office 😀 Short break to enjoy the long weekend and then back to the advisor matching process

2

0

64

Dylan Sam

@dylanjsam

3 years

Excited to be joining Amazon AWS AI @awscloud @AmazonScience as an Applied Scientist Intern this summer! I'm looking forward to (hopefully) being in California 🌞and working on topics in self-supervised learning

8

2

68

Dylan Sam

@dylanjsam

10 months

Check out our #NeurIPS2023 paper "Learning with Explanation Constraints" with my co-author @rpukdeee , which explains how explanations of model behavior can help us from a learning-theoretic perspective! () 🧵 (1/n)

2

16

59

Dylan Sam

@dylanjsam

2 years

1/ Here's a more thorough tweet thread on my paper: ("Losses over Labels: Weakly Supervised Learning via Direct Loss Construction"). I'll be presenting this at #AAAI23 on Sunday! 😀 Work w/ my advisor @zicokolter .

Losses over Labels: Weakly Supervised Learning via Direct Loss Construction

Owing to the prohibitive costs of generating large amounts of labeled data, programmatic weak supervision is a growing paradigm within machine learning. In this setting, users design heuristics...

arxiv.org

1

8

53

Dylan Sam

@dylanjsam

3 years

Finished a hybrid orientation at @mldcmu @CarnegieMellon !! It felt weird being back in lecture halls 😅 I'm excited for the upcoming semester of research and classes!

1

0

45

Dylan Sam

@dylanjsam

2 years

I'm attending #AAAI23 to present my work on a different perspective on weakly supervised learning: Please DM or email me if you are around and would like to chat! 😀

Losses over Labels: Weakly Supervised Learning via Direct Loss Construction

Owing to the prohibitive costs of generating large amounts of labeled data, programmatic weak supervision is a growing paradigm within machine learning. In this setting, users design heuristics...

arxiv.org

0

3

31

Dylan Sam

@dylanjsam

5 months

I'm at #ICLR2024 and will be presenting our work on understanding why prompt engineering generalizes so well! Come by our poster on 5/10 from 10:45am - 12:45pm 😀

Victor Akinwande

@aknvictor

1 year

The dominant paradigm in deep learning is to build larger-sized models. In recent work, we show that the statistical guarantees of such models do not necessarily suffer as a result of their sheer size. 🧵

2

26

113

0

4

31

Dylan Sam

@dylanjsam

3 years

1st semester of PhD done ✅ looking forward to heading home to recharge after a tough set of finals

0

30

Dylan Sam

@dylanjsam

1 year

I'm at #ICML2023 workshops - will be presenting some works on incorporating domain knowledge/weak supervision, identifying inequities in healthcare ⚖️, and understanding how prompting works! 🌴🍍 Thread below (1/n) 🧵👇

1

21

Dylan Sam

@dylanjsam

2 years

I'll be attending my first in-person conference #NeurIPS2022 to present my summer internship work () at the SSL workshop. Please reach out if you want to chat about SSL, weakly supervised learning, etc. Also down to watch the World Cup!

0

18

Dylan Sam

@dylanjsam

3 years

PhD student in CMU's MLD, and I make $36.7k/year

Gautam Goel

@gautamcgoel

3 years

To increase transparency around grad school stipends, retweet this tweet with your department, university, and annual stipend. I'll go first: I'm a PhD student in the Computing and Mathematical Sciences (CMS) department at Caltech, and I'm paid $36k/year. #StipendTransparency

81

189

1K

0

1

16

Dylan Sam

@dylanjsam

5 months

Check out this really interesting work led by @yewonbyun_ ! We propose a principled framework to identify unfair outcomes in existing decision-making systems ⚖️

Emily Byun

@yewonbyun_

5 months

Estimating notions of unfairness/inequity is hard as it requires that data captures all features that influenced decision-making. But what if it doesn't? In our work (), we answer this question w/ @dylanjsam @MichaelOberst @zacharylipton @brwilder

3

30

71

1

0

10

Dylan Sam

@dylanjsam

1 year

I'll be giving an oral presentation at the knowledge and logical reasoning workshop at 11am today on incorporating domain knowledge into Bayesian Neural Networks 🍌 (w/ @rpukdeee , @danielpjeong , @yewonbyun_ , @zicokolter ) (2/n)

1

2

7

Dylan Sam

@dylanjsam

3 years

Does anyone have suggestions for a first-time reviewer for ML conferences? 😀 What even makes a good review?

1

0

6

Dylan Sam

@dylanjsam

1 year

And our work on provably identifying inequity in resource allocation ⚖️, led by @yewonbyun_ (w/ @zacharylipton , @brwilder ) will be presented at the Interpretable Machine Learning for Healthcare workshop at 10:40am and 4:35pm (Ballroom C) (4/n)

1

6

Dylan Sam

@dylanjsam

3 years

Also, this is my first research internship, so I'm a bit excited and nervous to figure out this "industry" thing...

0

5

Dylan Sam

@dylanjsam

3 years

Come check out our #ICML2021 paper!!!

Stephen Bach

@stevebach

3 years

With Alessio, Cyrus, @dylanjsam99 , and Eli, we continue our investigation into theoretical guarantees for aggregating the votes of weak labelers. This time, we show how to provide guarantees in the multiclass setting. #ICML2021 Paper here:

0

12

0

4

Dylan Sam

@dylanjsam

1 year

And tomorrow, @aknvictor will be presenting our work (w/ @yidingjiang and @zicokolter ) on the shocking ability of prompts to generalize at the SODS workshop. The oral presentation is tomorrow at 2pm! (5/n)

0

4

Dylan Sam

@dylanjsam

10 months

This is joint work with co-author @rpukdee , @zicokolter , Nina Balcan, and Pradeep Ravikumar! @rpukdee are presenting (poster #1717 ) on Thurs @ 10:45am! (6/n)

0

1

Dylan Sam

@dylanjsam

3 years

@pliang279 we should do this for our papers to get maximum publicity 😂

1

0

2

Dylan Sam

@dylanjsam

10 months

There has been a recent line of research (XAI) on understanding black-box models through local explanations (e.g., feature importance). In practice, we may also have access to desirable explanations during training time, such as ignoring the background pixels of an image. (2/n)

1

0

1

Dylan Sam

@dylanjsam

2 years

4/ As a result, we name our method Losses over Labels (or LoL 😂).

1

0

2

Dylan Sam

@dylanjsam

2 years

@safranchik_e Yep, I agree! They seem to be much more informative than just sources of labels. I'll put together a tweet thread about this paper soon, it's long overdue...

0

2

Dylan Sam

@dylanjsam

2 years

6/6 Also, I've been working on some theory to provide more rigorous guarantees and clearer intuition on how this feature selection (or other forms of gradient regularization) helps. A paper to describe this is coming soon...

0

1

Dylan Sam

@dylanjsam

3 years

@Paragonadey @CarnegieMellon Go ahead!

3

0

1

Dylan Sam

@dylanjsam

10 months

We further analyze \emph{how} much of a reduction in sample complexity for a canonical class of input gradient explanations. (4/n)

1

0

1

Dylan Sam

@dylanjsam

3 years

@richardjchen Awesome, thanks!

0

1

Dylan Sam

@dylanjsam

3 years

@dkaushik96 @mldcmu @CarnegieMellon Maybe because you gave a presentation? 😄

0

1

Dylan Sam

@dylanjsam

3 years

@Paragonadey @CarnegieMellon Should be fixed now!

1

0

1

Dylan Sam

@dylanjsam

2 years

5/ We compare our method across tasks from the WRENCH benchmark and on the Animals with Attributes 2 dataset. We show that LoL improves upon existing WSL methods in many cases. Another key finding is that this gradient information leads to better performance on almost every task.

1

0

1

Dylan Sam

@dylanjsam

2 years

5/ The general intuition here is that this gradient information helps with feature selection, so gradient regularization (towards the mostly correct heuristics' gradients) leads to better performance. This is supported by experiments in settings with limited unlabeled data.

1

Dylan Sam

@dylanjsam

10 months

We can use this information to constrain our model to match our prior desirable explanation, and we show that this can benefit overall sample complexity. (3/n)

1

0

1

Dylan Sam

@dylanjsam

2 years

@jessicadai_ @berkeley_ai congrats jess!

0

1

Dylan Sam

@dylanjsam

2 years

3/ Instead, we transform the heuristics into losses that penalize differences between our model and the heuristics. This allows us to naturally incorporate more information, such as how the heuristics make their decisions (i.e., smoothed approximations of their gradients).

1

0

1

Dylan Sam

@dylanjsam

3 years

@jama1017 @CRAtweets @BrownCSDept Congrats Jiaju!!

0

1

Dylan Sam

@dylanjsam

4 years

@nishanthkumar23 @MITEECS @MIT_LISLab @MIT_CSAIL Congrats, man!!! Looking forward to reading all the great work you'll do

1

0

1

Dylan Sam

@dylanjsam

2 years

2/ In the standard weak supervision setting, users design noisy heuristics that are combined to form pseudolabels, which are used to train a downstream model. We question this pipeline; given that the heuristics provide all "label" information, why generate pseudolabels at all?

1

0

1

Dylan Sam

@dylanjsam

10 months

Finally, we provide a variational approach for this setting, which leads to better performance and better satisfaction of constraints given limited labeled data! (5/n)

1

0

1

Dylan Sam

@dylanjsam

3 years

@mobasiolu somehow i went from 0x to 6x a week 😀

1

0

1

Dylan Sam

@dylanjsam

1 year

And presenting a poster with @rpukdeee on our work on how prior knowledge about model behavior can help from a learning theoretic perspective: (3/n)

Learning with Explanation Constraints

As larger deep learning models are hard to interpret, there has been a recent focus on generating explanations of these black-box models. In contrast, we may have apriori explanations of how...

arxiv.org

1

0

1