Conceptron: a probabilistic deep one-class classification method

Erika Gardini, Andrea Cavalli, Sergio Decherchi

Keywords: unsupervised learning, one-class learning, anomaly detection, nystrom method, kernel methods

Abstract: One-class learning through deep architectures is a particularly challenging task; in this scenario the crasis of kernel methods and deep networks can represent a viable strategy to empower already effective methods. In this contribution we present Conceptron, a probabilistic and deep one-class classification method. The proposed algorithm is a hybridization of the Nystrom version of the Import Vector Domain Description (IVDD) to deep learning layers rendering the approach highly scalable (via batch stochastic gradient optimization) and automatically learning the underlying feature space. Further we modify the cost function to allow to get a Laplace distribution of the samples probabilities. Experiments on MNIST, CIFAR-10 and other benchmark datasets show that Conceptron (and/or variations) performs comparably or better with competing state-of-the-art methods with the additional capability of providing probabilities (through a logistic model) and avoiding any degeneracy in the training process.

One-sentence Summary: In this paper we introduce a one-class deep learning method, dubbed Conceptron, that bears state of the art prediction ability, explicit probability estimations and avoids degenerate solutions.