Anomaly Detection and Zero-Shot Learning with Error-Correcting Output Codes

Hieu Nguyen, SCOTT ZOCKOLL, Shen-Shyang Ho, MOHAMMED S. KHAN, LUCAS J. LAVALVA, and MATHEW
R. MARCHIANO

Zero-shot learning (ZSL) is a process by which a machine learns to recognize previously unseen objects. ZSL is a crucial step in artificial intelligence as a machine has to self-learn new concepts without a human teacher continuously providing labeled training data to the machine. Many approaches have been proposed using transfer learning of seen objects or knowledge transfer of seen data to a semantic embedded space to learn the new concept. We propose a new direction for ZSL in a more realistic incremental learning setting such that when a machine recognizes an unseen object as an anomaly (a deviation from all seen objects), it encodes the object as a new concept and integrates the new concept to the existing predictive model. In particular, we take an error-correcting output code (ECOC) driven approach in machine learning where class labels are represented by codewords from an error-correcting code that, when arranged as a matrix, has large row and column Hamming distances. Using multilabel classification, anomalies are detected and encoded as new concepts based on the output codewords of the unseen objects and their Hamming distances with respect to seen objects. Preliminary results on a set of datasets show the feasibility of our proposed approach.