For machines that are supposed to be getting more intelligent than any human alive, eating the whole web for breakfast, chatbot fatigue seems quite starkly --paradoxically -- real. (3/3)
English
Rama Vedantam
716 posts
Rama Vedantam
@rama_vedantam
AI Researcher | x-FAIR | https://t.co/GKsSzIxSjf
New York Katılım Aralık 2009
101 Takip Edilen905 Takipçiler
This is a huge contrast to people. The smartest people I have met get more and more interesting the more you talk to them, peeling layer after layer of insight, subtlety and personality. (2/3)
English
I'm pretty sure this has been said before, but IMHO the biggest problem with LLMs is that you can get them to say anything you want.
English
Hyperbolic Safety-Aware Vision-Language Models
arxiv.org/abs/2503.12127
Important problem. Nice idea. Cool embedding visualizations.
English
it is my great honour to be appointed as the Glen se Vries Professor of Health Statistics.
i have quickly written about this in my blog post: kyunghyuncho.me/glen-de-vries-…
NYU Courant@NYU_Courant
Kyunghyun Cho (@kchonyc), Professor of Computer Science and Data Science, has been named recipient of the Glen de Vries Chair for Health Statistics by the Courant Institute and New York University. Congratulations!
English
🛋️ Couch Philosophy
Reason has been seduced by a long history of fruitless and quarrelsome philosophical speculation.
Where should AI reasoning draw the line between useful abstraction and endless debate?
[5/6]
English
Deconstructing the mysteries of class conditional data 📚augmentation in ImageNet models. #NeurIPS2023
Meta-lesson: big data and big models can interact with each other in surprising, & interesting ways!
Led by @polkirichenko, collab with @andrewgwils @nyuniversity @AIatMeta
Polina Kirichenko@polkirichenko
Come to our poster at NeurIPS! neurips.cc/virtual/2023/p… W/ amazing co-authors @randall_balestr Mark Ibarhim @D_Bouchacourt @rama_vedantam @mamhamed @andrewgwils And check out Randall’s thread too! x.com/randall_balest… 9/9
English
At @NeurIPSConf next week. DMs are open.. ping me to catch up!!
English
Abstaining correctly is very important for open ended tasks such as VQA.
Checkout our work at #CVPR2023 led by the amazing @cdancette @marcus_rohrbach on this important topic.
Fantastic collaboration!
Corentin Dancette@cdancette
I am very excited to be at @CVPR in Vancouver this week to present the last work of my PhD, "Improving Selective VQA by Learning From Your Peers". We propose a method and benchmark to improve reliability of VQA models. See you there ! openaccess.thecvf.com/content/CVPR20…
English
Rama Vedantam retweetledi
Pending #CVPR2023 in June, we are pleased to share our 4 accepted papers
(4/4) "Improving Selective VQA by learning from your peers"
by @cdancette, S. Whitehead, R. Maheshwary, @rama_vedantam, S. Scherer, @endernewton, @quobbe, @marcus_rohrbach
Collab w/ @MetaAI
English
Flying on a one-way ticket in a long time.
See you in Bangalore ❤️
English
🔥 🔥 Our alternative to CLIP that captures visual-semantic hierarchies 🥳, is more efficient for smaller embeddings ✅, has geometric insights for multimodal learning 🤝🔥🔥
.. led by @kdexd (ex-intern, FAIR)
AK@_akhaliq
Hyperbolic Image-Text Representations abs: arxiv.org/abs/2304.09172
English
@BlackHC Yup, key is to do this with the right choice of basis. That is our technical contribution.
English
@rama_vedantam Thanks for explaining!
So you remove information from the embeddings that are confusing to the classifier because features that were never seen during training change the predictions in unintended ways?
English
Our paper on "Nullspace Occupancy" as a failure mode for distributional robustness in deep learning just got accepted into #iclr2023 [1/3]. Key Idea 👇
English
@BlackHC I don't think so. Intuitively it's more like IID inputs span a subspace of 2D and OOD inputs lie in the Nullspace of the IID subspace. Hence the name "Nullspace Occupancy"..
English
@rama_vedantam Is this the same as feature collapse in other works?
Different inputs get mapped to the same embeddings causing failures for the model to be able to differentiate them?
English