mk.
Portrait of Moshe Kimhi
HOW MUCH DETAIL IS ENOUGH?
A STUDY IN BOTTLENECKS

Reshape the bottleneck. Trace a signal.

Representation Bottleneck

01 / ABOUT · COMPUTER VISION & DEEP LEARNING

Moshe Kimhi

Learning more.
With less.

Research Scientist, Google Research · PhD researcher, Technion

I study neural networks under real constraints: less compute, fewer labels, and more possibility.

I’m a Research Scientist at Google Research and a PhD student in Computer Science at the Technion, advised by Chaim Baskin and Ehud Rivlin.

My research spans vision-language models, representation learning, semi-supervised learning, noisy labels, and robust AI systems. I hold BSc and MSc degrees in Electrical and Computer Engineering from the Technion. I teach deep learning and computer vision.

Have an interesting problem in mind? Let’s schedule a coffee ↗

A little context

02 / PERSPECTIVE

A modern
glassblower.

I see myself as a modern glassblower: finding the narrow neck, then reshaping it.

In neural networks, the bottleneck might be data, compute, memory, or the pipeline connecting them. I’m interested in identifying what limits a system, understanding why, and opening up room for it to do more.

I’m a strong believer in Jevons paradox ↗: greater efficiency can increase total demand. For me, making AI more efficient is a way to unlock new uses, broader access, and more ambitious questions.

That perspective connects my work on adaptive resolution, learning with imperfect data, quantization, and neural network robustness. The recurring question is simple: where does an intervention matter most?

03 / PUBLICATIONS

Find the constraint.
Expand what’s possible.

From visual tokens to training labels, I look for the small changes that make a large difference.

012026
Preprint · 2026

Sfumato Supervision: Learning Instance Segmentation with Noisy Labels

Audit 24,809 overlapping COCO–LVIS images, benchmark human and foundation-model contour noise, and break through the single-pixel boundary entropy floor via a directional Fréchet copula.

Authors & affiliations

Moshe Kimhi, Ehud Rivlin, Chaim Baskin

Kimhi & Rivlin: Technion · Baskin: Ben-Gurion University.

BibTeX
@article{kimhi2026sfumato,
 author = {Moshe Kimhi and Ehud Rivlin and Chaim Baskin},
 journal = {arXiv preprint arXiv:2406.10891},
 title = {{Sfumato Supervision: Learning Instance Segmentation with Noisy Labels}},
 url = {https://arxiv.org/abs/2406.10891},
 year = {2026}
}
022026
EMNLP Workshop on Document Intelligence and Understanding · 2026Best paper

Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs

Start with the big picture. Retrieve high-resolution evidence only where the question needs it.

Authors & affiliations

Nimrod Shabtay*, Moshe Kimhi*, Artem Spector, Sivan Haray, Ehud Rivlin, Chaim Baskin, Raja Giryes, Eli Schwartz

Shabtay: IBM Research, Tel Aviv University · Kimhi: IBM Research, Technion · Spector & Haray: IBM Research · Rivlin: Technion · Baskin: Ben-Gurion University · Giryes: Tel Aviv University · Schwartz: IBM Research. *Equal contribution.

BibTeX
@misc{shabtay2026lookmattershighresolutioncrops,
 archiveprefix = {arXiv},
 author = {Nimrod Shabtay and Moshe Kimhi and Artem Spector and Sivan Haray and Ehud Rivlin and Chaim Baskin and Raja Giryes and Eli Schwartz},
 eprint = {2603.16932},
 primaryclass = {cs.CV},
 title = {{Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs}},
 url = {https://arxiv.org/abs/2603.16932},
 year = {2026}
}
032026
ACL · 2026Oral presentation

CARES: Context-Aware Resolution Selector for VLMs

How much detail does a question really need? Select sufficient image resolution before invoking the VLM.

Authors & affiliations

Moshe Kimhi*, Nimrod Shabtay*, Raja Giryes, Chaim Baskin†, Eli Schwartz†

Kimhi: IBM Research, Technion · Shabtay: IBM Research, Tel Aviv University · Giryes: Tel Aviv University · Baskin: Ben-Gurion University · Schwartz: IBM Research. *Equal contribution; †Equal supervision.

BibTeX
@inproceedings{kimhi-etal-2026-cares,
 address = {San Diego, California, United States},
 author = {Kimhi, Moshe  and
Shabtay, Nimrod  and
Giryes, Raja  and
Baskin, Chaim  and
Schwartz, Eli},
 booktitle = {Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)},
 doi = {10.18653/v1/2026.acl-long.102},
 editor = {Liakata, Maria  and
Moreira, Viviane P.  and
Zhang, Jiajun  and
Jurgens, David},
 isbn = {979-8-89176-390-6},
 month = {July},
 pages = {2243--2256},
 publisher = {Association for Computational Linguistics},
 title = {{{CARES}: Context-Aware Resolution Selector for {VLM}s}},
 url = {https://aclanthology.org/2026.acl-long.102/},
 year = {2026}
}
042026
TMLR · 2026Featured certificate · Top 10%

Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips

A handful of critical sign bits can break a neural network. Finding them reveals where protection matters most.

Authors & affiliations

Ido Galil*, Moshe Kimhi*, Ran El-Yaniv

Galil: NVIDIA · Kimhi: Technion · El-Yaniv: Technion, NVIDIA. *Equal contribution.

BibTeX
@article{galil2026maximal,
 author = {Ido Galil and Moshe Kimhi and Ran El-Yaniv},
 journal = {Transactions on Machine Learning Research},
 title = {{Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips}},
 url = {https://openreview.net/forum?id=kN1s53X3zl},
 year = {2026}
}
052026
Preprint · 2026

CXR-Retrieve: Compositional Text-to-Image Retrieval in Chest Radiography

Retrieve chest radiographs that satisfy compositional clinical queries.

Authors & affiliations

Tomer Erez, Moshe Kimhi, Chaim Baskin, Ehud Rivlin

Erez, Kimhi & Rivlin: Technion · Baskin: Ben-Gurion University.

BibTeX
@misc{erez2026cxrretrievecompositionaltexttoimageretrieval,
 archiveprefix = {arXiv},
 author = {Tomer Erez and Moshe Kimhi and Chaim Baskin and Ehud Rivlin},
 eprint = {2607.27779},
 primaryclass = {cs.CV},
 title = {{CXR-Retrieve: Compositional Text-to-Image Retrieval in Chest Radiography}},
 url = {https://arxiv.org/abs/2607.27779},
 year = {2026}
}
062026
WACV · 2026

R³: Reconstruction, Raw, and Rain: Deraining Directly in the Bayer Domain

Remove rain before the image processing pipeline discards useful raw sensor information.

Authors & affiliations

Nate Rothschild, Moshe Kimhi, Avi Mendelson, Chaim Baskin

Rothschild, Kimhi & Mendelson: Technion · Baskin: Ben-Gurion University.

BibTeX
@inproceedings{Rothschild_2026_WACV,
 author = {Rothschild, Nate and Kimhi, Moshe and Mendelson, Avi and Baskin, Chaim},
 booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
 month = {March},
 pages = {5313-5321},
 title = {{R3: Reconstruction, Raw, and Rain: Deraining Directly in the Bayer Domain}},
 year = {2026}
}
072025
Preprint · 2025

WAVECLIP: Wavelet Tokenization for Adaptive-Resolution CLIP

Read images coarse to fine, reusing computation and refining only when necessary.

Authors & affiliations

Moshe Kimhi*, Erez Koifman*, Ehud Rivlin, Eli Schwartz†, Chaim Baskin†

Kimhi: Technion, IBM Research · Koifman & Rivlin: Technion · Schwartz: IBM Research · Baskin: Ben-Gurion University. *Equal contribution; †Equal supervision.

BibTeX
@misc{kimhi2025waveclipwavelettokenizationadaptiveresolution,
 archiveprefix = {arXiv},
 author = {Moshe Kimhi and Erez Koifman and Ehud Rivlin and Eli Schwartz and Chaim Baskin},
 eprint = {2509.21153},
 primaryclass = {cs.CV},
 title = {{WAVECLIP: Wavelet Tokenization for Adaptive-Resolution CLIP}},
 url = {https://arxiv.org/abs/2509.21153},
 year = {2025}
}
082025
WACV · 2025

Robot Instance Segmentation with Few Annotations for Grasping

Learn from scene changes and unlabeled images to reduce annotation demands for robotic grasping.

Authors & affiliations

Moshe Kimhi, David Vainshtein, Chaim Baskin, Dotan Di Castro

Kimhi: Technion, Bosch Center for AI · Vainshtein & Di Castro: Bosch Center for AI · Baskin: Ben-Gurion University.

BibTeX
@inproceedings{Kimhi_2025_WACV,
 author = {Kimhi, Moshe and Vainshtein, David and Baskin, Chaim and Di Castro, Dotan},
 booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
 month = {February},
 pages = {7928-7938},
 title = {{Robot Instance Segmentation with Few Annotations for Grasping}},
 year = {2025}
}
092024
TMLR · 2024

Semi-Supervised Semantic Segmentation via Marginal Contextual Information

Use neighboring pixels together to refine pseudo labels and learn more from unlabeled images.

Authors & affiliations

Moshe Kimhi, Shai Kimhi, Evgenii Zheltonozhskii, Or Litany, Chaim Baskin

All authors: Technion · Litany additionally: NVIDIA AI.

BibTeX
@article{kimhi2024semisupervised,
 author = {Moshe Kimhi and Shai Kimhi and Evgenii Zheltonozhskii and Or Litany and Chaim Baskin},
 journal = {Transactions on Machine Learning Research},
 title = {{Semi-Supervised Semantic Segmentation via Marginal Contextual Information}},
 url = {https://openreview.net/forum?id=i5yKW1pmjW},
 year = {2024}
}
102024
NeurIPS ENLSP Workshop · PMLR 262 · 2024

Hysteresis Activation Function for Efficient Inference

Revisit the activation function to address dying ReLUs with minimal inference complexity.

Authors & affiliations

Moshe Kimhi, Idan Kashani, Chaim Baskin, Avi Mendelson

Kimhi, Kashani & Mendelson: Technion · Baskin: Ben-Gurion University. Author order follows the published PMLR record.

BibTeX
@inproceedings{pmlr-v262-kimhi24a,
 author = {Kimhi, Moshe and Kashani, Idan and Baskin, Chaim and Mendelson, Avi},
 booktitle = {Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop},
 editor = {Rezagholizadeh, Mehdi and Passban, Peyman and Samiee, Soheila and Partovi Nia, Vahid and Cheng, Yu and Deng, Yue and Liu, Qun and Chen, Boxing},
 month = {14 Dec},
 pages = {414--422},
 pdf = {https://raw.githubusercontent.com/mlresearch/v262/main/assets/kimhi24a/kimhi24a.pdf},
 publisher = {PMLR},
 series = {Proceedings of Machine Learning Research},
 title = {{Hysteresis Activation Function for Efficient Inference}},
 url = {https://proceedings.mlr.press/v262/kimhi24a.html},
 volume = {262},
 year = {2024}
}
112024
Preprint · 2024

Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods

Refine context embeddings to combine in-context learning with parameter-efficient adaptation.

Authors & affiliations

Tsachi Blau, Moshe Kimhi, Yonatan Belinkov, Alexander Bronstein, Chaim Baskin

Blau, Kimhi, Belinkov & Bronstein: Technion · Baskin: Ben-Gurion University.

BibTeX
@misc{blau2024contextawareprompttuningadvancing,
 archiveprefix = {arXiv},
 author = {Tsachi Blau and Moshe Kimhi and Yonatan Belinkov and Alexander Bronstein and Chaim Baskin},
 eprint = {2410.17222},
 primaryclass = {cs.CL},
 title = {{Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods}},
 url = {https://arxiv.org/abs/2410.17222},
 year = {2024}
}
122024
Mathematics 12(12), 1810 · 2024

AMED: Automatic Mixed-Precision Quantization for Edge Devices

Choose mixed precision with the target hardware in mind.

Authors & affiliations

Moshe Kimhi, Tal Rozen, Avi Mendelson, Chaim Baskin

All authors: Computer Science Department, Technion.

BibTeX
@article{Kimhi_2024,
 author = {Kimhi, Moshe and Rozen, Tal and Mendelson, Avi and Baskin, Chaim},
 doi = {10.3390/math12121810},
 issn = {2227-7390},
 journal = {Mathematics},
 month = {June},
 number = {12},
 pages = {1810},
 publisher = {MDPI AG},
 title = {{AMED: Automatic Mixed-Precision Quantization for Edge Devices}},
 url = {http://dx.doi.org/10.3390/math12121810},
 volume = {12},
 year = {2024}
}
132022
Mathematics 10(21), 4107 · 2022

Bimodal-Distributed Binarized Neural Networks

Shape weight distributions to preserve useful representations under binarization.

Authors & affiliations

Tal Rozen, Moshe Kimhi, Brian Chmiel, Avi Mendelson, Chaim Baskin

All authors: Technion · Chmiel additionally: Habana Labs, an Intel company.

BibTeX
@article{rozen2022bimodal,
 author = {Tal Rozen and Moshe Kimhi and Brian Chmiel and Avi Mendelson and Chaim Baskin},
 doi = {10.3390/math10214107},
 journal = {Mathematics},
 number = {21},
 pages = {4107},
 title = {{Bimodal-Distributed Binarized Neural Networks}},
 url = {https://www.mdpi.com/2227-7390/10/21/4107},
 volume = {10},
 year = {2022}
}

04 / TALKS

Ideas, in conversation.

Mobileye · July 13

Efficient Image Encoding for Vision-Language Models

Google Research Conference · October 13

Memory in Perception Encoders for Autonomous Driving

05 / SET A MEETING

Let’s talk AI & vision.

I consult and teach on AI and computer vision, helping teams turn research ideas into practical systems. Get in touch for consulting, tailored teaching sessions, or a speaking event.

Schedule a meeting ↗

For consulting, a teaching session, or a speaking event, tell me a little about what you have in mind.

Email me directly: moshekimhi91@gmail.com ↗