Lifelong Context Recognition via Online Deep Feature Clustering

Context recognition for lifelong learning (L2) agents is an open-ended problem whereby aggregate features in an environment are utilized to signal the active context in which the agent is operating. The ability to recognize context is necessary in L2 agents to engage modulatory signals to account for significant changes in the input state space associated with a given context or task, such as altering learning dynamics or shifting attention to more relevant features. Context recognition is itself an L2 problem due to the ever-increasing number of distinct contexts that an agent might encounter, requiring incrementally learning novel contexts while prescribing them to supervised task labels when available. This paper demonstrates an algorithm based on near clustering of deep-extracted features with adaptive resonance theory methods that satisfies these requirements on the behalf of an embodied L2 agent in a computer vision environment. The strength of this algorithm lies in its flexibility, being capable of online incremental learning in supervised, realistic semi-supervised, and unsupervised scenarios while demonstrating continual learning in its own right.