The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Equivariant Priors for Compressed Sensing with Arash Behboodi - #584
15 snips
Jul 25, 2022 
In this discussion, Arash Behboodi, a machine learning researcher at Qualcomm Technologies, dives deep into his groundbreaking paper on equivariant generative models for compressed sensing. He explains how these models can recover signals with unknown orientations, offering theoretical recovery guarantees. The conversation touches on evolving VAE architectures, the challenges of signal recovery in wireless communication, and the exciting applications of his work in fields like cryo-electron microscopy. Additionally, they explore innovative strategies in quantization-aware training and temporal causal identifiability.
AI Snips
Chapters
Transcript
Episode notes
Cryo-Electron Microscopy Challenge
- Cryo-electron microscopy takes noisy pictures of biomolecules in unknown orientations.
- Multiple pictures have different, unaligned orientations.
Equivariant Priors
- Jointly recover signal and orientation using equivariant generative models.
- These models embed orientation information, eliminating separate orientation discovery.
Equivariance in Generative Models
- Equivariant models have been used in supervised learning, but their application to generative models is relatively new.
- Arash Behboodi's team developed a novel equivariant VAE.