A computational model of learning to count in a multimodal, interactive environment

AbstractWhen learning to count, children actively engage with a variety of counting tasks and observe demonstrations by more knowledgeable others. We investigate how a single neural network-based agent, situated in a multimodal learning environment, can learn from observing such demonstrations to perform multiple number tasks such as counting temporally and spatially distributed objects, and a variant of the ‘give-N’ task. We find that i. the agent can learn different tasks that require counting, ii. learning progresses in similar stages for different tasks, iii. sequential learning of subtasks aids learning of the full counting task, and iv. a mechanism for updating memory when each object is counted emerges from learning the task. The work relies on generic deep learning processes in widely used neural network modules rather than mechanisms specialized for mathematics learning, and provides an architecture in which a sense of number emerges from learning several different number related tasks.

Return to previous page