Moshe Kimhi
About Me
I am a **PhD student** in Computer Science at the **Technion**, advised by [Ehud Rivlin](https://research.google/people/ehud-rivlin/) and [Chaim Baskin](https://chaimb.cs.technion.ac.il/). My research focuses on **learning under constraints**—tackling semi-supervised learning, noisy labels, limited compute, and robust AI systems. I'm particularly interested in **representation learning, vision-language models, and efficient neural networks**. I hold a BSc and MSc in Electrical and Computer Engineering from the Technion. Beyond research, I serve as a reviewer for top-tier venues (ICCV, ICML, WACV, TMLR) and co-teach computer vision and deep learning.Featured Publications
Predicts minimal sufficient input resolution for image-query pairs in VLMs, reducing compute by up to 80% while preserving performance.
Deep neural lesion (DNL)—a data-free method identifying critical sign bits. Flipping just 2 bits in ResNet-50 reduces ImageNet accuracy by 99.8%.
RISE
Semi-supervised learning for robotic grasping with temporal consistency. Achieves 86.37 AP50 on ARMBench—20% improvement over prior work.
S4MC
Enhances pseudo-labeling in semi-supervised segmentation by leveraging spatial correlations. 1.29 mIoU improvement on PASCAL VOC 12.
HeLU
Proposes HeLU to address the dying ReLU problem with minimal complexity. Achieves competitive performance across diverse architectures.
AMED
Mixed-precision quantization using Markov Decision Processes for optimal bitwidth allocation on specific hardware.
BD-BNN
Bimodal-distributed binarization method using kurtosis regularization. Outperforms prior BNN schemes on CIFAR-10 and ImageNet.
Preprints
Preprint
Comprehensive benchmark of label noise effects in instance segmentation. Introduces COCO-N, CityScapes-N, and VIPER-N datasets.
Preprint
Combines in-context learning and prompt tuning with adversarial methods for superior classification accuracy.
Recognition & Service
Reviewer: ICCV, ICML, WACV, TMLR
Teaching: Co-instructor for CS 236781 (Computer Vision & Deep Learning), Technion