This academic article presents a compressed sensing (CS) setup for the reconstruction of Generative Sparse-Latent (GSL) signals, which are generated by a neural network excited by a sparse-latent input signal. The proposed reconstruction algorithm uses a gradient-based search to induce sparsity, providing a good reconstruction performance despite the inherent non-convexity of GSL signals. The article also introduces a measure to understand the degree of non-linearity in the mapping between the measurement and the sparse-latent signal, which helps in performing controlled simulation studies. This research contributes to the novel application of sparse-latent signals for generative models in a CS setup.
Publication date: 24 Oct 2023
Project Page: Not Given
Paper: https://arxiv.org/pdf/2310.15119
