/DeepLS

Primary LanguageJupyter Notebook

DeepLS

DeepLS is a novel unsupervised deep learning approach designed for online separation of sparse signal components. This method leverages the low-rank and sparse priors inherent in datasets for training purposes. It utilizes a U-Net-based model, structured similarly to an hourglass, which efficiently encodes and decodes sparse components. During training, the model employs a loss function based on a combination of nuclear and $\ell_1$ norms, mirroring the objective function of Robust Principal Component Analysis (RPCA). This approach encourages the model to discern and isolate the sparse components within the input data.

Notably, once trained, the DeepLS model gains the capacity to extract learned sparse components from a diverse range of inputs, not limited to those exhibiting low-rank characteristics. This capability eliminates the need for retraining the model for varying backgrounds, leading to computationally efficient online separation of target signals that share features with the learned sparse components from the training dataset.

Applications and code samples