Jonathan Kahana

Jonathan Kahana is a Researcher in the Computer Science department at the Hebrew University of Jerusalem. His research spans Machine Learning and Computer Vision, focusing on Weight Space Learning, Representation Learning, and Zero-Shot Model Search. He develops methods for probing neural network weights to extract information, including ProbeGen and Spectral DeTuning.

His recent work includes mapping model weights into shared embedding spaces (ProbeX), recovering pre-fine-tuning weights of generative models, and improving zero-shot labeling with distribution priors. He contributes to open-source implementations, such as the ProbeGen GitHub repository.

Research trends from his publications emphasize:

• Weight space analysis (ProbeGen, DSiRe)
• Model retrieval and classification (ProbeLog, Model Atlas)
• Disentanglement and invariance (Contrastive Objective, Red PANDA)
• Efficiency in probing (30-1,000x FLOPs reduction in ProbeGen)

His work has been accepted at top-tier conferences including ICML, ICLR, and ECCV, with arXiv preprints covering topics like dataset size recovery and model tree analysis.