(from lapidow & @ebonawitz.bsky.social's awesome 2023 explore-exploit paper)

The present work thus serves as a ‘case study’ of sorts. It yields concrete discoveries about real-world size, and it also validates a broadly applicable tool for psychology and neuroscience. We hope it catches on!

Though we manipulated real-world size, you could generate anagrams of happy faces and sad faces, tools and non-tools, or animate and inanimate objects, overcoming low-level confounds associated with such stimuli. Our approach is perfectly general.

Overall, our work confronts the longstanding challenge of disentangling high-level properties from their lower-level covariates. We found that, once you do so, most (but not all) of the relevant effects remain.

(Never fear, though: As we say in our paper, that last result is consistent with the original work, which suggested that mid-level features — the sort preserved in ‘texform’ stimuli — may well explain these search advantages.)

Finally, visual search. Previous work shows targets are easier to find when they differ from distractors in their real-world size. However, in our experiments with anagrams, this was not the case (even though we easily replicated this effect with ordinary, non-anagram images).

Next, aesthetic preferences. People think real-world large objects look better when displayed large, and vice versa for small objects. Our experiments show that this is true with anagrams too!

First, the “real-world size Stroop effect”. If you have to say which of two images is larger (on the screen, not in real life), it’s easier if displayed size is congruent with real-world size. We found this to be true even when the images were perfect anagrams of one another!

Then, we placed these images in classic experiments on real-world size, to see if observed effects arise even under such highly controlled conditions.

(Spoiler: Most of these effects *did* arise with anagrams, confirming that real-world size per se drives many of these effects!)

We generated images using this technique (see examples). Each pair differs in real-world size but are otherwise identical* in lower-level features, because they’re the same image down to the last pixel.

(*avg orientation, aspect-ratio, etc, may still vary. ask me about this!)

This challenge may seem insurmountable. But maybe it isn’t! To overcome it, we used a new technique from Geng et al. called “visual anagrams”, which allows you to generate images whose interpretations vary as a function of orientation.

Take real-world size. Tons of cool work shows that it’s encoded automatically, drives aesthetic judgments, and organizes neural responses. But there’s an interpretive challenge: Real-world size covaries with other features that may cause these effects independently.

The problem: We often study “high-level” image features (animacy, emotion, real-world size) and find cool effects. But high-level properties covary with lower-level features, like shape or spatial frequency. So what seem like high-level effects may have low-level explanations.