EE Seminar: DBI: Deep Bispectrum Inversion
https://tau-ac-il.zoom.us/j/84026490379
Electrical Engineering Systems ZOOM Seminar
Speaker: Keren Cohen
M.Sc. student under the supervision of Prof. Tamir Bendory
Wednesday, 17th September 2025, at 15:00
DBI: Deep Bispectrum Inversion
Abstract
This work addresses the problem of signal reconstruction from the bispectrum - also known as bispectrum inversion. The bispectrum preserves phase information and exhibits shift invariance. These properties make it especially useful in tasks like multi-reference alignment (MRA), where multiple noisy and shifted observations of the same signal are available. A prominent real-world example of such a setting is the reconstruction of 3D molecular structures in cryogenic electron microscopy (cryo-EM), for which MRA serves as a one-dimensional abstraction.
This study addresses both the homogeneous setting, where the objective is to recover a single signal from its bispectrum, and the more challenging heterogeneous setting, in which a collection of distinct signals must be recovered from a mixture of their bispectra. The computational hardness of the problem increases substantially in the heterogeneous case, as the number of distinct underlying signals grows.
To recover a signal from its bispectrum, we investigate the potential of deep learning to handle more general cases, and propose a learning-based solution to this challenging inverse problem. The proposed model is based on supervised learning and incorporates convolutional layers, with optional integration of transformer blocks. We conducted extensive experiments involving the recovery of up to three distinct signals, which demonstrated reliable reconstruction performance in the homogeneous case while achieving a computationally efficient solution. The method lays a promising foundation for future exploration of the heterogeneous scenario.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום בצ'ט של שם מלא + מספר ת.ז.
