Abstract
Humans have the ability to rapidly understand rich combinatorial conceptsfrom limited data. Here we investigate this ability in the context of auditorysignals, which have been evolved in a cultural transmission experiment to studythe emergence of combinatorial structure in language. We propose aneuro-symbolic generative model which combines the strengths of previousapproaches to concept learning. Our model performs fast inference drawing onneural network methods, while still retaining the interpretability andgeneralization from limited data seen in structured generative approaches. Thismodel outperforms a purely neural network-based approach on classification asevaluated against both ground truth and human experimental classificationpreferences, and produces superior reproductions of observed signals as well.Our results demonstrate the power of flexible combined neural-symbolicarchitectures for human-like generalization in raw perceptual domains andoffers a step towards developing precise computational models of inductivebiases in language evolution.