Abstract
Recent approaches to human concept learning have successfully combined thepower of symbolic, infinitely productive rule systems and statistical learningto explain our ability to learn new concepts from just a few examples. The aimof most of these studies is to reveal the underlying language structuring theserepresentations and providing a general substrate for thought. However,describing a model of thought that is fixed once trained is against theextensive literature that shows how experience shapes concept learning. Here,we ask about the plasticity of these symbolic descriptive languages. We performa concept learning experiment that demonstrates that humans can change veryrapidly the repertoire of symbols they use to identify concepts, by compilingexpressions which are frequently used into new symbols of the language. Thepattern of concept learning times is accurately described by a Bayesian agentthat rationally updates the probability of compiling a new expression accordingto how useful it has been to compress concepts so far. By portraying theLanguage of Thought as a flexible system of rules, we also highlight thedifficulties to pin it down empirically.