From Continuous Observations to Symbolic Concepts: A Discrimination-Based Strategy for Grounded Concept Learning

Autonomous agents perceive the world through streams of continuous sensori-motor data. Yet, in order to reason and communicate about their environment, agents need to be able to distill meaningful concepts from their raw observations. Most current approaches that bridge between the continuous and symbolic domain are using deep learning techniques. While these approaches often achieve high levels of accuracy, they rely on large amounts of training data, and the resulting models lack transparency, generality, and adaptivity. In this paper, we introduce a novel methodology for grounded concept learning. In a tutor-learner scenario, the method allows an agent to construct a conceptual system in which meaningful concepts are formed by discriminative combinations of prototypical values on human-interpretable feature channels. We evaluate our approach on the CLEVR dataset, using features that are either simulated or extracted using computer vision techniques. Through a range of experiments, we show that our method allows for incremental learning, needs few data points, and that the resulting concepts are general enough to be applied to previously unseen objects and can be combined compositionally. These properties make the approach well-suited to be used in robotic agents as the module that maps from continuous sensory input to grounded, symbolic concepts that can then be used for higher-level reasoning tasks.