A deep learning account of how language affects thought

How can words shape meaning? Shared labels highlight commonalities between concepts whereas contrasting labels make differences apparent. To address such findings, we propose a deep learning account that spans perception to decision (i.e. labelling). The model takes photographs as input, transforms them to semantic representations through computations that parallel the ventral visual stream, and finally determines the appropriate linguistic label. The underlying theory is that minimising error on two prediction tasks (predicting the meaning and label of a stimulus) requires a compromise in the network's semantic representations. Thus, differences in label use, whether across languages or levels of expertise, manifest in differences in the semantic representations that support label discrimination. We confirm these predictions in simulations involving fine-grained and coarse-grained labels. We hope these and allied efforts which model perception, semantics, and labelling at scale will advance developmental and neurocomputational accounts of concept and language learning.