Fig 6. Three-step approach to obtaining prototype representations.

A A large-scale image search query provided cue category examples. For every cue category (i.e., trial), we performed a separate image search query returning multiple hundreds of photographs specific to every category cue seen by the human participants. B Describing categories in model representations. In a next step, we took all retrieved images and presented them to the single-image models (also serving as a baseline for the sequential models). Like this, we obtained an activation pattern for every image. C Computing the prototype representation. Together all category examples formed a category manifold. To obtain a prototype representation, we simply found the geometric mean across the activation patterns evoked by the different category exemplars. Note that only a subset of exemplars was used to obtain the final prototype.