@omrisuissa.com
Postdoc @ Brown DSI VP R&D @ ClearMash π¬ Passionate about high-fidelity numerical representations of reality, aligned with human perception. https://omri.alphaxiv.io/ #nlp #multimodality #retrieval #hci #multi-agent
Stop training VLMs to just name objectsβteach them concepts. π CLEAR GLASS (with the MAGIC dataset) uses a grouped contrastive loss to make abstraction emerge, beating CLIP & HierarCaps on abstract concept recognition. Read: arxiv.org/pdf/2509.12771
#AI #VLM #Multimodal οΏΌ
At the biu conference
I had the opportunity to give a talk at BIU on "Enriching Multimodality Embedding using Cross-modality Supervision." I explored how one modality can enhance another leading to a better alignment with human perspectives.
Grateful for the engaging discussions and the chance to share our work.
Yesterday at Brown University, I lectured on retrieval-augmented generation (RAG). I shared insights into its workings, real-world applications, and limitations. Iβm grateful to Brown University for hosting me and to everyone who participated.
CVPR reviews are out π₯³
Screw itβ¦
I'm impressed with the new Profile feature from @alphaxiv.org! It would be even better if they included more structured sections, such as teaching experience, awards, grants, and others.
omri.alphaxiv.io
Patents counts as legal? π love this domain
arxiv.org/abs/2411.08526
Just submitted a paper to cvpr 2025 on the topic. This is my main research area these days. Would love to be added π
MT is where I started my machine learning journey, love this domain
dl.acm.org/doi/abs/10.1...
πββοΈπ
πββοΈ love AI & culture. There are still many untapped opportunities for cultural preservation and investigation in the AI domain.
asistdl.onlinelibrary.wiley.com/doi/abs/10.1...
πββοΈ I worked on question answering on tables and graphs, love this topic
dl.acm.org/doi/abs/10.1...
πββοΈ I work on question answering on knowledge graphs, love this topic
content.iospress.com/articles/sem...
Close source AI risks: same as open sourcing AI risks - but without the chance to fix and validate them independently.
Iβm sorry for your loss π
Data science joke:
What would you say to an ML engineer who was unable to successfully train a model?
(answer in the first comment)
A modest yet highly satisfying gamification milestone.
π Honored to present "Introduction to Efficient LLMs" at Brown Universityβs DSI.
A big thank you to the DSI team and all the engaged students and faculty for your insightful questions during and after the talk π
#AI #MachineLearning #LLMs #Efficiency #BrownUniversity
m.youtube.com/watch?v=J8Ya...
CVPR requires all submissions to remain anonymous π€«π
We appear to be experiencing the βlaw of diminishing returnsβ in this area. While scaling remains effective and will continue to do so, the incremental value we gain is progressively decreasing.
This challenge extends beyond AI research; it also affects question-answering systems in enterprises, chatbots, and similar use cases. The lack of tacit knowledge (unwritten) is a fundamental barrier to replacing humans in any type of work
Interesting, thank you!