Ph.D. direct-track student at Tel Aviv University (Geometric Deep Learning Lab) · Lead, Deep Learning at Dtect Vision.
I study the foundations of deep learning — spectral bias, optimization dynamics, and generalization. My central question is how neural networks learn at different frequencies, and what that means for real-world learning problems. I treat learned representations as signals, characterize how training dynamics absorb and amplify spectral content, and use that lens to design principled methods.
The frequency-domain view connects problems that look unrelated. Two of the applications I focus on are detection of AI-generated audio, where the high-frequency residuals networks struggle to learn become a discriminative signal, and continual learning and catastrophic forgetting, where spectral analysis of representation drift exposes what is lost during training and what can be preserved.
At Dtect Vision I lead the deep learning group, translating these insights into production-grade detection systems. Academically, I am advised in the Geometric Deep Learning Lab, working on spectral-bias-aware optimization, frequency-guided representations, and the broader theory of how networks learn.
Prior: B.Sc. in Applied Mathematics and M.Sc. in Electrical Engineering (signal processing track), both at Tel Aviv University. I moved into the direct Ph.D. track after completing my M.Sc. thesis work.
Neural networks preferentially learn low frequencies, which both leaves high-frequency artifacts in generated audio and leaves those artifacts under-exploited by common detectors. SONAR is a dual-path framework that learns data-driven SRM-style filters to isolate high-frequency residuals and uses a Jensen–Shannon contrastive objective to pull real content/noise pairs together and push fake pairs apart in latent space, yielding strong generalization to unseen generators.
More preprints in preparation — on frequency-aware optimization and generalization in detection models.
Foundations of deep learning: spectral bias, optimization dynamics, generalization, with applications to synthetic-media forensics.
Signal processing track. Complementary coursework in random signals & noise, signal processing, and control theory (GPA 98).
Coursework in probability, statistics, optimization, stochastic processes, numerical methods, and machine learning.
How neural networks differentially absorb low- vs. high-frequency structure across training, and why the spectrum a model can represent is the spectrum it tends to learn.
How optimizer choice, parameterization, and loss geometry shape spectral trajectories during training, and how those trajectories predict downstream generalization.
Spectral signatures of catastrophic forgetting: which frequency bands of a learned representation drift first, and what spectrally-aware interventions preserve plasticity without sacrificing stability.
Frequency-residual methods for detection of AI-generated audio and image content, with an emphasis on cross-generator generalization and robustness under deployment shift.
PyTorch · Lightning · Torch DDP · Python / NumPy / Pandas · Docker · GCP · CNNs · encoders · generative models · loss design · ablations & reproducibility.
Email · Idonithid@gmail.com
Scholar · Ido Nitzan Hidekel
LinkedIn · ido-nitzan-hidekel
GitHub · @idonithid
Happy to discuss projects on frequency-domain analysis of neural networks, deepfake detection, or applied detection problems grounded in signal processing. Reach out by email.