Cognitively Inspired Learning of Incremental Drifting Concepts

Humans continually expand their learned knowledge to new domains and learn new concepts without any interference with past learned experiences. In contrast, machine learning models perform poorly in a continual learning setting, where input data distribution changes over time. Inspired by the nervous system learning mechanisms, we develop a computational model that enables a deep neural network to learn new concepts and expand its learned knowledge to new domains incrementally in a continual learning setting. We rely on the Parallel Distributed Processing theory to encode abstract concepts in an embedding space in terms of a multimodal distribution. This embedding space is modeled by internal data representations in a hidden network layer. We also leverage the Complementary Learning Systems theory to equip the model with a memory mechanism to overcome catastrophic forgetting through implementing pseudo-rehearsal. Our model can generate pseudo-data points for experience replay and accumulate new experiences to past learned experiences without causing cross-task interference. Humans continually abstract concept classes from their input sensory data to build semantic descriptions, and then update and expand these concepts as more experiences are accumulated Widmer & Kubat (1996), and use them to express their ideas and communicate with each other Gennari et al. (1989); Lake et al. (2015). For example, "cat" and "dog" are one of the first concept classes that many children learn to identify. Most humans expand these concepts as concept drift occurs, e.g., incorporating many atypical dog breeds into the "dog" concept, and also incrementally learn new concept classes, e.g. "horse" and "sheep," as they acquire more experiences. Although this concept learning procedure occurs continually in humans, continual and incremental learning of concept classes remains a major challenge in artificial intelligence (AI). AI models are usually trained on a fixed number of classes and the data distribution is assumed to be stationary during model execution.