Subhankar Roy

484 posts

Subhankar Roy banner
Subhankar Roy

Subhankar Roy

@sroy907

Tenure track Assistant Professor at University of Bergamo.

Katılım Nisan 2010
486 Takip Edilen272 Takipçiler
Subhankar Roy
Subhankar Roy@sroy907·
@CSProfKGD Can we get a meme for a reviewer who promised to raise the score from BR in the initial review, but didn't show up after the rebuttal? :)
English
1
0
1
160
Kosta Derpanis
Kosta Derpanis@CSProfKGD·
After the first round of reviews, your paper looks like a shoo-in for acceptance … then Reviewer 2 shows up with a new argument to reject 😡
Kosta Derpanis tweet media
English
1
0
17
2.8K
Tomer Galanti
Tomer Galanti@GalantiTomer·
A bit late to the game. Excited to share that 1329/1330 of my ICLR submissions were accepted! 😇 Even though one paper was unjustly desk-rejected I am hopeful that it will get into ICML. Wish me luck 😁
English
6
2
106
19.4K
Kosta Derpanis
Kosta Derpanis@CSProfKGD·
#KostasThoughts: The #ICLR2026 review process has been disappointing. This is not on the PCs. It reflects our current environment where the scale of reviewing across the various conferences may have outgrown the bandwidth of the community (eg shortcuts via LLM generated reviews) and where some feel pressured enough to lose sight of basic ethics (eg reports of authors contacting reviewers to change their score or much worse if social media reports are to be believed). Perhaps it is time to rethink our conference culture. Share your thoughts in thread.
English
9
5
76
22.5K
Subhankar Roy retweetledi
Max Zhdanov
Max Zhdanov@maxxxzdn·
The variance across ICLR scores is wild, OpenAI should decrease the temperature
English
8
44
617
34.3K
Johan
Johan@Adityapandeydev·
how can i prevent my macbook keyboard from becoming like this?
Johan tweet media
English
1.4K
66
4.3K
995K
Subhankar Roy retweetledi
vik
vik@vikhyatk·
@GFaang97609 no i never run ablations, because i believe god has a plan
English
8
19
192
14.7K
Subhankar Roy
Subhankar Roy@sroy907·
Having a paper, which was accepted by the AC, rejected by the PCs is really atrocious. This can damage the morale of young researchers, especially PhD students. Quota-based acceptance should be stated in the cfp so that the authors can calibrate their expectations. #neurips25
English
0
1
6
470
abhishek
abhishek@abhi1thakur·
I no longer have doubts
abhishek tweet media
English
213
340
11.8K
1.2M
Subhankar Roy
Subhankar Roy@sroy907·
We submitted a paper on semantic segmentation to a CV conference that shall not be named. Reviewer's comment, "The approach involves extracting high-quality domain knowledge from expert-written documents" Dear reviewer, kindly choose a better LLM next time.
English
0
0
6
198
Subhankar Roy
Subhankar Roy@sroy907·
This year 11 researchers from University of Oxford have received ERC starting grants! To put things into perspective to those who don't know, an institute is fortunate to have just one accepted. Success rate is ~12%. Congratulations to all the recipients.
Subhankar Roy tweet media
English
1
0
1
133
Subhankar Roy
Subhankar Roy@sroy907·
@giffmana In that era of unsupervised deep learning, there was this clip order prediction pretext task for video representation learning.
English
0
0
1
223
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
An aux task for LLM to predict the order of next chunk of tokens. Haha this is so 2018! In computer vision this kind of tasks was all the rage for "self-supervised" learning! This specific one was called "jigsaw". There was also "relative patch location" and "rotation prediction". All follow the same pattern: you artificially mess with the image, and the model needs to find out what you did. Actually, this was my first paper with @__kolesnikov__ and @XiaohuaZhai , where we basically said: "wait a minute, all of these aux tasks are kind of a big mess" (while also setting new sota using them, so we are allowed to say that lol) It's interesting because i see more and more previous vision things becoming relevant to language modeling lately, i love it!
Lucas Beyer (bl16) tweet media
tokenbender@tokenbender

this is the paper. just found it out in the wild and thought of this post.

English
17
27
375
94.9K
Subhankar Roy
Subhankar Roy@sroy907·
@abeirami IMO, safety in AI algorithms is induced through patching and setting explicit guardrails. True safety can never be achieved without conscientiousness. We are nowhere close to that.
English
1
0
1
336
Ahmad Beirami
Ahmad Beirami@abeirami·
I am not sure how such a pause helps. My thoughts as "Not an AI Professor": 1. AI safety is not a theoretical discipline. We will only make the models safe and secure by building them in the first place, observing the pitfalls, and iterating over fixing them. 2. We are not anywhere close to reaching superintelligence and at least another technological leap is needed. So, we are not even equipped with solving the safety problems of superintelligence if we paused model training now.
David Krueger 🦥 ⏸️ ⏹️ ⏪@DavidSKrueger

I'm an AI professor. Superintelligence could arrive any year, and we are not prepared.

English
10
4
98
21.1K
Illia
Illia@Zieeett·
Nano Banana: Make Image Daytime and Isometric (Building Only)
Illia tweet mediaIllia tweet media
English
75
229
5.6K
898.9K