Self-Communicating Deep Reinforcement Learning Agents Develop External Number Representations

Symbolic numbers are a remarkable product of human cultural development. The developmental process involved the creation and progressive refinement material representational tools, such as notched tallies, knotted strings, counting boards. In this paper, we introduce computational framework that allows investigation how representations might support number processing in deep reinforcement learning scenario. framework, agents can use an external, discrete state to communicate information solve simple numerical estimation task. We find different perceptual constraints result emergent representations, whose specific characteristics facilitate communication numbers.