Brain-Inspired Concept Networks: Learning Concepts from Cluttered Scenes

Abstract

It's unclear how our brain's concepts emerge sequentially, and how the brain abstracts and generalizes each concept internally. The artificial intelligence field has seen the rise of model-based methods, where, starting from a predefined set of tasks, a programmer handcrafts a model for the set. Such a machine is incapable of generating and using any concept beyond the handcrafted model. Inspired by the anatomical connection patterns in the cerebral cortex, the authors introduce concept networks as an embodiment of the more general class of brain-inspired developmental networks. Such a network acquires concepts as actions through autonomous, incremental, and optimal self-wiring and adaptation according to its learning and practicing experience. Recursively, a concept network generates the current actions, which serve as its dynamic concepts to direct its next internal operation, which then generates the next actions. As the network learns and practices in an open-ended manner, its concepts aren't restricted by a handcrafted world model.