Unsupervised feature selection (UFS) is a key step in image processing and pattern recognition, but existing approaches have a trade-off: model-based methods depend on manually tuned hyperparameters, while deep learning methods typically require large training sets. In a new arXiv preprint, Jingjing Liu and colleagues propose a network that aims to get the best of both worlds.

Named MDUFS-Net, the method incorporates the l2,0-norm into a classic PCA framework and then 'unfolds' the iterative ADMM optimization algorithm into an end-to-end trainable network. This lets the model learn its parameters automatically rather than relying on hand-set values. The authors also introduce a manifold optimization strategy to improve numerical robustness.

An enhanced version, MDUFS-Net+, replaces the l2,0-norm with the l0-norm to obtain more discriminative feature representations. In extensive experiments, both variants outperform state-of-the-art UFS methods and show promising clustering performance. The code is publicly available.