We can quickly tell whether (and in what direction) a stack of precarious objects will fall. But studying "intuitive physics" in this way requires asking observers about physics in the first place. Here we asked how such physical relationships are (implicitly) encoded into memory. We used a deceptively simple block reconstruction task: observers briefly viewed randomly generated block towers, then rebuilt them from memory, all without ever being told anything about stability or physics. When we later simulated the influence of gravity on both the original and reconstructed towers, we discovered a surprising bias: observers remembered unstable towers as more stable than they initially were, even when their reconstructions were otherwise quite accurate. This suggests that visual memory represents not only what objects are present and where they are, but also how they physically relate to one another.

In order to do research on how we see, remember, and think about faces, you need to use face stimuli. But how do you source faces for your research? Leveraging recent advances in deep learning and big data, we created hyperrealistic generative models of face impressions, allowing for the creation and transformation of face photos along 34 dimensions of psychological interest. In addition, the models can predict how people will judge face photos on all these dimensions, including "trustworthiness", "attractiveness", and "dominance". We are currently building an API for researchers across myriad fields to make use of these models in their own work — especially those working to combat discrimination in the wider world.

Facial stereotypes are not only harmful, but often inaccurate. Even so, they reliably drive people's judgments and behavior. How do such biases propagate despite weak evidence? We used a novel implementation of the method of iterated learning to reveal that people's stereotypes of facial trustworthiness are "sticky". Participants were tasked with learning a relationship between how "trustworthy" a face looked to how "trustworthily" the person behaved in an economic game, and were then tested on their knowledge of that relationship. However, the answers given at test time became the "training" for the next participant in the chain of reproduction. Within only ten generations, participants reproduced a familiar steretotype: more trustworthy-looking faces behaved more trustworthily, even when there was no relationship between looks and behavior to begin with. This work demonstrates that visual stereotypes are powerful, easily spread, and may be surprisingly difficult to root out.

Faces are arguably the most important visual stimuli in our social world. How are they encoded and represented in the mind? Our work explores the notion of 'mental defaults' in face space — prioritized regions in the multidimensional space that "pull" encoded representations towards them. Using the method of serial reproduction, we showed that white observers' default face representation is white: even a recognizably Black face is transformed into a recognizably white face within only ten reproductions.