Gagan Jain
@gaganjain1582
Followers
219
Following
1,002
Media
15
Statuses
48
@GoogleDeepMind | @IITBombay | Opinions my own
Joined November 2020
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
TEMU
• 430613 Tweets
最大5,000円報酬
• 226528 Tweets
領空侵犯
• 147426 Tweets
Sven
• 114661 Tweets
Arlington
• 91599 Tweets
Delpino
• 47250 Tweets
シロネン
• 46621 Tweets
Finale
• 34177 Tweets
Scout
• 27731 Tweets
Nintendo Direct
• 24296 Tweets
線状降水帯
• 23780 Tweets
Indie World
• 23769 Tweets
DUNK SMN EP4
• 22489 Tweets
PHUWIN SUMMERNIGHT EP4
• 22045 Tweets
Switchの後継機種
• 21372 Tweets
ソフトメーカーラインナップ
• 21129 Tweets
Rayo Vallecano
• 20470 Tweets
D-1 TO CHANYEOL SOLO
• 16697 Tweets
James Rodríguez
• 15186 Tweets
#NationalDogDay
• 11937 Tweets
伊藤さん
• 10258 Tweets
Pinned Tweet
Your models about to get a MoNE-y makeover! 💸 We're thrilled to unveil MoNE (Mixture of Nested Experts), a game-changer for efficiency & scalability. Get ready to level up! 🚀
Google presents Mixture of Nested Experts
Adaptive Processing of Visual Tokens
The visual medium (images and videos) naturally contains a large amount of information redundancy, thereby providing a great opportunity for leveraging efficiency in processing. While Vision
4
77
379
1
14
52
👀A sneak-peek at our recent work accepted at
#ECCV2024
@eccvconf
!
📜:
Want a scalable and robust foundational model that also saves computational costs? Say hello to LookupViT, our information compression idea leading to sub-quadratic attention!
(1/10)
First conference submission ☑️
First conference acceptance ✅
#eccv2024
@eccvconf
Congrats Sujoy
@rajatkoner
@jainprateek_
🎉
More details soon :D
6
2
73
3
24
140
First conference submission ☑️
First conference acceptance ✅
#eccv2024
@eccvconf
Congrats Sujoy
@rajatkoner
@jainprateek_
🎉
More details soon :D
6
2
73
#NeurIPS
Reviewer: Do X, Y, Z
We perform these (somewhat intensive) exps, while also justifying why some of these weren't even needed given paper results
Reviewer (< a day to discussion deadline): Won't increase score, I think more extensive evals/discussion needed
-_-
1
0
6
We're humbled by the early positive reception and shares for our MoNE paper! 🙏 A huge shoutout to my amazing collaborators and mentors for their incredible contributions -
@paul_sujoy_
Nidhi Hegde
@adityakusupati
@jainprateek_
@NagraniArsha
Shyamal Buch, Anurag Arnab ✨
1
1
5
Some
#NeurIPS
reviewers need these things:
1. A reminder that their task is to evaluate potential scientific advancements in the field
2. An open and unbiased mind
3. An alarm clock
4. Magnifying Glasses
0
0
5
So many interesting keynotes, today's twitter feed is giving me major FOMO. Next year's
@RL_Conference
on my bucket list now 😤
Great keynote by David Silver, arguing that we need to re-focus on RL to get out of the LLM Valley
@RL_Conference
11
60
432
0
0
4
Not all tokens provide new information, yet self-attention based models do not exploit this redundancy.
This impacts their scalability, and calls for newer techniques that compress information, and can be adaptively used at inference based on task complexity.
(2/10)
1
0
4
How? We sample compressed tokens (M) from the original input token pool (N) (called lookup tokens).
These two sets of tokens are then fed to a series of LookupViT blocks, which process and bidirectionally exchange information to enhance these tokens.
(4/10)
1
0
4
@paul_sujoy_
@adityakusupati
@jainprateek_
@NagraniArsha
Stay tuned for more updates on MoNE & its exciting applications across different domains! In an era where efficient computation is paramount, we believe MoNE's approach to conditional computation has the potential to reshape the future of large-scale ML models. 🚀
1
0
4
Key Idea: We maintain two sets of tokens that exchange information amongst themselves -
1⃣ Fine-grained but many, that are cheaply processed
2⃣ Coarse but few, that are heavily processed
(3/10)
1
0
4
Curious to learn more? Dive into the technical details in our arXiv paper: 🤓
1
0
4
[Bangalore House Rent Experience]
I'll happily pay a broker once than use
@nobrokercom
's scam plans ever again. Operating without a shred of ethic, their so called relationship managers will ignore all complaints & get their "brokerage" from your sec. deposit when you vacate.
1
2
4
The outcome?
We beat vanilla ViT and some state-of-the-art compression techniques on standard benchmarks, demonstrate scalability with increasing image resolution, and show applicability to tasks like video classification and image captioning!
(6/10)
1
0
4
Several open directions of research, including auto-selection of compression token set based on task complexity - an important step towards conditional computation.
Would love to chat and exchange more ideas on efficiency of large models.
Looking forward to Milan 🇮🇹:)
(9/10)
1
0
3
We've already seen promising results on image & video datasets – MoNE boosts efficiency without sacrificing performance. 👀
1
0
3
Cherry on top, FlexiViT style training allows adaptive choice of the number of compressed tokens at inference time from a single model, with minimal accuracy loss!
Extensive ablations and visualizations in the paper investigate the why’s and how’s about our framework.
(8/10)
1
0
3
MoNE's secret sauce? 🪄 A clever combo of nested experts & adaptive routing, allowing it to smartly allocate resources based on input complexity.
1
0
3
And, and, and, we show significantly superior robustness than other baselines in the presence of adversarial perturbations! We further show that the feature corruption with increasing severity levels is lower for LookupViT than vanilla ViT.
(7/10)
1
0
3
Work with awesome collaborators & mentors
@GoogleDeepMind
@LMU_Muenchen
:
@rajatkoner
^
@paul_sujoy_
@jainprateek_
@vtresp
!
Extra pumped, for this is my 1st A* conf paper. Extremely grateful to
@ManishGuptaMG1
for the constant support!
^ denotes equal contribution
(10/10)
0
0
3
Information processing involves three key steps:
Information gathering: Information flow from lookup to compressed tokens
Representation refinement: Compressed tokens undergo heavy computation
Global Context Infusion: Information flow from compressed to lookup tokens
(5/10)
1
0
3
No more wasted compute! MoNE tackles redundant or simple info with lightweight experts, keeping your models lean & mean. 💪
1
0
3
#NeurIPS
reviews done and dusted (just move on), here's something to look forward to at
#NeurIPS2024
. Also, a chance to interact with the ever amazing (and only slightly scary)
@nikitasaxena02
✨
(1/n) 📢 Thrilled to be part of the organizing team of the WiML Workshop at
#NeurIPS2024
!
🧠 Full CFP:
Submission Deadline: September 10, 2024
@WiMLworkshop
@NeurIPSConf
@WiCVworkshop
@WiNLPWorkshop
#WiML
#NeurIPS
#machinelearning
#WiML2024
#WiML2024
2
11
44
0
0
2
@fooobar
@paul_sujoy_
@jainprateek_
Thank you so much for your kind words, Gaurav :)) Grateful to be here, all thanks to you!
0
0
1
1
0
1
@PavloMolchanov
@ccccrs_0908
@srv_m
@gLeHeinrich
@yin_hongxu
@VITAGroupUT
@jankautz
Awesome work. Congrats! It was great talking to you during ICML :))
0
0
1
Love this
Young scientists regularly ask me for career advice. Academia or industry? Big company or startup? US or Europe? Good scientists in AI disciplines are fortunate to have many choices. But choosing can be stressful. I always give the same advice. 1/10
21
250
1K
0
0
1
@amit_sethi
All in for note taking if the subject is mathematical in nature. When learning something for the first time, I have often found note taking to be the difference between "yeah that makes sense" and "oh I see why it makes sense".
0
0
1