Research

Our lab develops new deep learning tools for signal processing and computational imaging. In particular, we focus on deep learning theory, few-shot learning, deep-optics, and solving inverse problems.  

Broadly speaking, the lab research is at the intersection between the fields of signal and image processing, and machine learning. More specifically, our interests include, but are not limited to, deep learning, sparse representations, low dimensional signal modeling, computational imaging, compressed sensing, and inverse problems.

Some of the lab recent works include (i) a new fidelity term for solving inverse problems with an advanced strategy to use GANs as priors by training them on the target data (ii) a neural network for 3D meshes; (iii) a correction filter that allows using super-resolution neural networks with any down-sampling kernel (CVPR 2020 award nominee); (iv) deep learning-based optical design for all-in-focus imaging and single-lens depth reconstruction (first place in the OSA optical element of the future grand student challenge); (v) new few-shot learning tools for image classification, object detection, multi-label images, interpretable weakly supervised detection and self-supervised based fine-class classification; (vi) a deep neural network that replaces the conventional image signal processing (ISP) pipeline; (vii) novel adversarial robustness and detection approaches; (viii) theoretical analysis of neural network, e.g.,  by studying their margin-based generalization, smoothness impact on expressivity and generalization and using tools from signal processing.

White Structure