Towards a more flexible language of thought: Bayesian grammar updates after each concept exposure

Abstract

Recent approaches to human concept learning have successfully combined the power of symbolic, infinitely productive rule systems and statistical learning to explain our ability to learn new concepts from just a few examples. The aim of most of these studies is to reveal the underlying language structuring these representations and providing a general substrate for thought. However, describing a model of thought that is fixed once trained is against the extensive literature that shows how experience shapes concept learning. Here, we ask about the plasticity of these symbolic descriptive languages. We perform a concept learning experiment that demonstrates that humans can change very rapidly the repertoire of symbols they use to identify concepts, by compiling expressions that are frequently used into new symbols of the language. The pattern of concept learning times is accurately described by a Bayesian agent that rationally updates the probability of compiling a new expression according to how useful it has been to compress concepts so far. By portraying the language of thought as a flexible system of rules, we also highlight the difficulties to pin it down empirically.

A logical framework to study concept-learning biases in the presence of multiple explanations

“…We conjecture that most first-order determinants of subjective concept difficulty will also hold in a relative manner in our dual-concept setup, such as the MDL bias (for less extreme cases than evaluated in this work) (Feldman, 2003) and the transfer learning hierarchical structure bias (Tano et al, 2020). Importantly, our experimental setup also allows to directly test second-order subjective difficulty effects (e.g.…”

“…Although different methods have been used to present the concept to the participant (e.g. all elements at the same time (Tano et al, 2020 ; Kemp, 2012 ) or small sets of elements presented in series Piantadosi et al, 2016 ), to the best of our knowledge all previous category-learning studies have attempted to evaluate a single concept at a time. Here, we present a controlled logical setting to evaluate the relative difficulty of two concepts presented at the same time and under the same experimental conditions, and the framework could be generalized to more concepts straightforwardly.…”

“…p 1 ∧ p 2 and p 3 ∨ p 4 ) using the same set of examples has several experimental advantages over separately presenting a set of examples consistent with rule A and then a set of examples consistent with rule B . Some of the advantages are: When comparing the relative difficulty of learning A and B in the same participant, presenting the examples separately makes it hard to overcome transfer effects that cause subjective difficulty to depend on the history of concepts learnt previously in the task, and cause different relative difficulties if A is learnt before B compared to B being learnt before A (see for example Tano et al, 2020 ). The experimenter could compare learning times for A and B across participants, but for reasonably hard rules there are very large idiosyncratic differences in learning difficulties which greatly increases the variance of learning times (see for example Feldman, 2000 ), and also the experimenter cannot normalize the past history of each participant before the experiment.…”

Seeking new information with old questions: Children and adults reuse and remix concepts from prior questions

“…When making inferences, people incorporate the output of previously-formed inferences (Dasgupta et al, 2018;Gershman & Goodman, 2014, see also Banavar & Bornstein, 2023). When generating hypotheses, people build on previously-endorsed hypotheses, rather than starting from scratch (Gelpi et al, 2020;Herbst et al, 2017;Tano et al, 2020;Zhao et al, 2024, see also Ellis et al, 2023;Tian et al, 2020). In other words, when innovating, people "remix" their own and others' prior ideas.…”

Compositionality Under Time Pressure

“…However, a limitation of our experimental design is that while participants were introduced to the primitives during the tutorial, they were not trained on them. Although it is known that people can acquire new primitives automatically and without explicit training (Tano, Romano, Sigman, Salles, & Figueira, 2020), the capacity for compositionality has also been linked to hippocampal replay (Schwartenbeck et al, 2021;Kurth-Nelson et al, 2023), which may require longer timescales of training. Future studies should investigate how prior knowledge about primitives effects the capability to deploy them in a compositional fashion.…”