Proceedings of the Annual Meeting of the Cognitive Science Society

When learning to count, children actively engage with a varietyof counting tasks and observe demonstrations by more knowl-edgeable others. We investigate how a single neural network-based agent, situated in a multimodal learning environment,can learn from observing such demonstrations to perform mul-tiple number tasks such as counting temporally and spatiallydistributed objects, and a variant of the give-N task. We findthat 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 task ofcounting spatially distributed objects, and iv. a mechanism forupdating memory when each object is counted emerges fromlearning the task. The work relies on generic deep learningprocesses in widely used neural network modules rather thanmechanisms specialized for mathematics learning, and pro-vides an architecture in which aspects of a sense of numberemerge from learning several different number related tasks.