Itamar Avitan

I use deep neural networks to understand the computations behind biological vision and behavior. I test whether our methods can really tell those models apart.

About

I want to understand the computations behind biological vision and the cognition built on it, and I use deep neural networks to look for them. These networks already predict human vision and behavior remarkably well; the harder problem now is telling them apart. My recent work puts our methods to that test: with model-recovery simulations at realistic signal-to-noise, I ask what current alignment metrics can reveal about the model behind the data, where they fail, and how we could measure better.

I’m drawn to statistical and Bayesian modeling, and to experiments designed so that competing models disagree. Much of my work is hands-on deep learning: aligning many vision networks to brain responses and to millions of behavioral judgments and running large simulations on a GPU cluster. I work with Dr. Tal Golan in the Brains and Machines Lab at Ben-Gurion University of the Negev.

Research

NeurIPS 2025

When the Best-Fitting Model Isn’t the Right One (Research page)

Is the network that predicts human judgments best also the right one? We fitted 20 networks to 4.5 million human odd-one-out judgments, let each generate judgments, and asked whether the comparison would pick it out. It picked the right one less than 80% of the time: far above the one in twenty of guessing, but wrong more than one time in five, even with millions of simulated trials. That is a result for the task, models, noise calibration and mappings we tested, not a claim that every flexible evaluation fails.

  1. generator

  2. judgments

  3. candidates

  4. ranking

Now

  • Teaching assistant again for Introduction to Computation and Cognition, and for Academic Writing from 2026/27.

News

  • Gave a flash talk, “System identification perspective on representational alignment” (with Maya Zach and Tal Golan), at Neuro-AI-Talks (NEAT) 2026 in Osnabrück, Germany.

  • Completed the Embodied Brain Technology Practicum, a joint Brown University–Ben-Gurion University program held at Brown, where our team of five built and pitched EarBetter.

  • Presented our paper with Tal Golan as a poster at the Conference on Neural Information Processing Systems (NeurIPS) 2025 in San Diego.

Older news (4)

  • Released the code and data for our NeurIPS 2025 paper.

  • Posted the arXiv preprint of the same paper, after its acceptance to the NeurIPS 2025 main track.

  • Gave a contributed talk and presented a poster at the Cognitive Computational Neuroscience (CCN) 2025 conference in Amsterdam, on a preliminary version of what became our NeurIPS 2025 paper.

  • Started my PhD in the Brains and Machines Lab at Ben-Gurion University of the Negev, on a Lachish Fellowship for Outstanding Doctoral Students.