Representations of geometric shapes have syntactic structure w/ @maxencepajot.bsky.social and @standehaene.bsky.social is out & open-access in JEP:General doi.org/10.1037/xge0.... For an overview, see thread below!
Thus, humans (and so far only humans) are predisposed to encode a variety of data in syntactic structures—an ability that extends well beyond natural language! 8/8
Preprint: osf.io/preprints/ps...
We also tested convolutional and transformer networks on the three tasks by examining the pairwise distances between the embeddings for familiarization vs test shapes. The networks solve the tasks, but they show no evidence of syntax in their representations. 7/8
Inspired by syntactic movement, Experiment 3 shows that shape fragments are easier to put back together the higher up in the tree they are split: Going one level down in the tree imposes a constant processing cost → Compelling, parametric evidence for hierarchy. 6/8
Inspired by clausal segmentation, Experiment 2 shows that subparts are easier to recognize when they belong to the same subtree than when they span multiple subtrees → More evidence of structure and indirect evidence of hierarchy. 5/8
Inspired by structural ambiguity, Experiment 1 shows that the same shape can be assigned distinct representations based on how a preceding drawing animation organizes it → Shape representations are structured. 4/8
Current study: We provide more direct evidence for the LoT format of geometric-shape representations by moving from complexity measures to constituency tests, building on previous work on the geometry LoT by @mathiassablemeyer.bsky.social and colleagues. 3/8
Background: A growing body of work points to the existence of multiple languages of thought (LoT), each with its own set of primitives and grammar. But most evidence supporting LoT representations relies on an indirect measure: stimulus compression as afforded by the hypothesized languages. 2/8
Preprint alert (link below)! @maxencepajot.bsky.social, @standehaene.bsky.social, and I show that human adults, but not convolutional or transformer networks, encode geometric shapes in hierarchically structured representations. TL;DR: Geometric-shape representations have internal syntax! 1/8