Research
My current work at Dolby focuses on multimodal video understanding. My PhD research covered computational imaging, robust machine learning and learning theory.
Multimodal Understanding of Long-Form Video
Models that combine video, audio and text to retrieve moments, answer questions, localize actions and curate highlights in long videos.
Other Dolby Projects
HDR for Dolby Vision 2, and real-time active speaker detection for mobile devices and smart glasses.
Inverse Imaging with Deep Untrained Neural Network Priors
Untrained deep networks as image priors for compressed sensing, phase retrieval and HDR imaging, with convergence guarantees.
Adversarially Robust Learning via Entropic Regularization
ATENT: an entropy-regularized loss for training neural networks that are robust to adversarial perturbations.
Provable algorithms for machine learning
Gradient descent and alternating minimization for shallow ResNets, with linear convergence guarantees.
Compressive phase retrieval under image priors
CoPRAM and Block CoPRAM: fast, sample-efficient phase retrieval using sparsity and structured sparsity.
Fourier ptychography for super-resolution under image priors
Sub-diffraction imaging via Fourier ptychography using sparsity and low-rank priors.
Older Projects
Course and side projects from graduate school: CGI image forensics, topic modeling with NMF and sparse PCA, and compressive video segmentation.