Deep learning researcher · PhD, Technion (2026)
My recent work builds calibrated diagnostics for language models: effective depth, the shuffle gap and invertible query-key coupling, with two first-author papers at ACML 2026. Alongside, I keep classical structure inside the learning loop: a differentiable Randers-Finsler eikonal solver for wildfire spread (SGP 2026, Computer Graphics Forum) and physiologically constrained models that estimate VO₂ from smartwatches and from 3D human pose in video (ACM ICMI 2025, Apple). PhD advised by Alex Bronstein and Reuven Cohen.
Open to research scientist and applied scientist roles in LLM understanding, physics-informed AI and sensing.
An identity-initialized, invertible coupling applied before scoring that adds 0.3% parameters and composes with Differential Attention, with significant gains at 150M and 455M parameters.
A calibrated scalar for how much of a transformer's depth is distinct computation: fifteen of sixteen decoder-only models fall below the orthogonal-update reference, and matched controls trace the gap to correlated residual updates.
A GPU fast-sweeping solver made exactly differentiable through implicit differentiation, used to recover anisotropic metrics from sparse arrival times and learn spread models on real wildfire perimeters.
Full VO₂ trajectories from smartwatch and chest-strap data alone: a physiologically constrained ODE with a neural Kalman filter for heart rate, and a one-second-calibration VO₂ model at about 11% MAPE.
Calibration-free diagnostics that audit a model instead of trusting its headline scores: how much of its depth is distinct computation, whether a personalized model conditions on the user at all, and where pre-scoring query-key geometry helps attention.
Classical solvers kept inside the learning loop. Implicit differentiation through fast sweeping turns anisotropic Eikonal distance fields into trainable models of wildfire spread.
Instantaneous VO₂ and cardiovascular load estimated from smartwatches, chest straps and monocular 3D pose in video, with neural ODE and Kalman structure and four datasets collected in the field.
QUIC traces rendered as image sequences and multi-operator cellular links as multivariate series, from a 100,000-trace public dataset to real-time network selection for teleoperated vehicles.