Markov-Lipschitz Deep Learning

In this paper, we propose a novel framework, called Markov-Lipschitz deep learning (MLDL), for manifold learning and data generation. It enhances layer-wise transformations of neural networks by optimizing isometry (geometry-preserving) and stability of the mappings. A prior constraint of locally isometric smoothness (LIS) is imposed across-layers and encoded via a Markov random field (MRF)-Gibbs distribution, and consequently layer-wise vector transformations are replaced by LIS-constrained metric homeomorphisms. These lead to the best possible solutions as measured by locally geometric distortion and locally bi-Lipschitz continuity. Extensive experiments, comparisons and ablation study demonstrate significant advantages of MLDL for manifold learning and data generation.