Open-World Classification with Bayesian Gaussian Mixture Models

Methods for solving classification tasks often assume a data generating process with stable structure that remains fixed during both training and inference. However, autonomous agents deployed in real-world environments often perform classification in situations where the data generating process is dynamic and the ontology of classes is only partially known. Such tasks are known as openworld classification (OWC). We present open-world mixture modeling (OMM), a framework for OWC using Bayesian Gaussian mixture models. With only slight modifications to the standard Bayesian variational inference algorithm, we are able to detect and model novel classes as they appear in a data stream, while maintaining and updating the classes learned during training. Empirical evaluations reveal that the method reliably detects novel classes with performance similar to a supervised classifier trained on labeled samples of the novel classes.

Authors

• Justin Clarke University of Massachusetts
• Przemyslaw Grabowicz
• David Jensen

Keywords

• Open-World Learning
• Multi-Class Classification
• Continual Learning