Deep Inductive and Scalable Subspace Clustering via Nonlocal Contrastive Self-Distillation

Deep subspace clustering has demonstrated remarkable results by leveraging the nonlinear subspace assumption. However, it often encounters challenges in terms of computational cost and memory footprint in dealing with large-scale data due to its traditional single-batch training strategy. To address this issue, this paper proposes a deep subspace clustering framework that is regularized by nonlocal contrastive self-distillation, enabling a Deep Inductive and Scalable Subspace Clustering (DISSC) algorithm. In particular, our framework incorporates two subspace learning modules, namely subspace learning based on self-expression model and inductive subspace clustering. These modules generate affinities from different perspectives by extracting intermediate features from two augmentations of the input data using a weight-sharing neural network. By integrating the concept of self-distillation, our framework effectively exploits the clustering-friendly knowledge contained in these two affinities through a novel nonlocal contrastive prediction task, employing an empirical yet effective threshold. This allows the framework to facilitate complementary knowledge mining and scalability without compromising clustering performance. With an alternate branch that bypasses the self-expression computation, our framework can infer subspace membership of the out-of-sample data through the predicted soft labels, eliminating the need for ad-hoc postprocessing. In addition, the self-expression matrix computed using mini-batch data benefits from the distilled knowledge obtained from the inductive subspace clustering module, enabling our framework to scale to data of arbitrary size. Experiments conducted on large-scale MNIST, Fashion-MINST, STL-10, CIFAR-10 and Stanford Online Products datasets validate the superiority of the proposed DISSC algorithm over state-of-the-art subspace clustering methods.