Conference on Cognitive Computational Neuroscience

Compositional generalisation is the ability to decompose and re-combine predictive factors (or rules) to navigate new contexts. Here, we replicate and extend a recent study of this ability in humans by Dekker et al., (2022). We compare participants’ ability to generalise a mapping of nonspatial features (colour, shape, texture) to new 3D spatial locations using a compositional rule, under training curricula that block or interleave the features. We replicate the advantage of blocked curricula for compositional generalisation, finding that interleaving tends to instead promote exemplar-based strategies.