A consistent and flexible framework for deep matrix factorizations
作者：
Highlights：
• Two new loss functions for deep matrix factorizations are introduced.
• Our loss functions are weighted sums of layer-based errors.
• A general convergent optimization framework is presented to tackle our formulations.
• Several priors are incorporated, e.g., sparsity, nonnegativity and minimum-volume.
• Experiments on synthetic and real data validate the effectiveness of our framework.
摘要
•Two new loss functions for deep matrix factorizations are introduced.•Our loss functions are weighted sums of layer-based errors.•A general convergent optimization framework is presented to tackle our formulations.•Several priors are incorporated, e.g., sparsity, nonnegativity and minimum-volume.•Experiments on synthetic and real data validate the effectiveness of our framework.
论文关键词：Deep matrix factorization,Loss functions,Constrained optimization,First-order methods,Hyperspectral unmixing
论文评审过程：Received 19 June 2022, Revised 2 September 2022, Accepted 4 October 2022, Available online 12 October 2022, Version of Record 12 October 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.109102
