Proceedings of the Annual Meeting of the Cognitive Science Society

Recent advances in Deep Learning (DL) and ReinforcementLearning (RL) make it possible to train neural network agentswith raw, first-person visual perception to execute language-like instructions in 3D simulated worlds. Here, we inves-tigate the application of such deep RL agents as cognitivemodels, specifically as models of infant word learning. Wefirst develop a simple neural network-based language learningagent, trained via policy-gradient methods, which can inter-pret single-word instructions in a simulated 3D world. Tak-ing inspiration from experimental paradigms in developmentalpsychology, we run various controlled simulations with the ar-tificial agent, exploring the conditions in which established hu-man biases and learning effects emerge, and propose a novelmethod for visualising and interpreting semantic representa-tions in the agent. The results highlight the potential util-ity, and some limitations, of applying state-of-the-art learningagents and simulated environments to model human cognition.