SFU MOCAD Seminar: Simone Brugiapaglia

Since its inception in the early 2000s, compressive sensing has become a well-established paradigm for efficient signal recovery, with applications ranging from medical imaging to scientific computing. More recently, data-driven reconstruction methods based on deep neural networks have attracted considerable attention and shown great promise as an alternative approach. In this talk, we will review recent progress in signal reconstruction techniques that combine principles from compressive sensing and deep learning. First, we will discuss recent advances in generative compressive sensing, where the traditional sparsity prior is replaced by the assumption that the signal to be reconstructed lies in the range of a deep generative neural network. Second, we will explore deep greedy unfolding, which involves designing deep neural network architectures by "unrolling" the iterations of a sparse recovery algorithm onto the layers of a trainable neural network. In both cases, we will present numerical results in tandem with theoretical guarantees.