In machine learning, the learner is assumed to be rational, seeking the highest probability, lowest cost, account of the data. This may not be achievable without a tremendous amount of data. Meanwhile, the problem of overfitting remains a formidable challenge: the best hypothesis from one set of data may not generalize well to another. Herbert Simon observed that human learning and decision-making often do not strive for optimal solutions but merely solutions that are good enough ("satisficing"). Language learning provides the most compelling demonstration. Almost all linguistic rules have exceptions but are nevertheless good enough to generalize. A "Tolerance Principle" of learning by satisficing provides a precise and parameter-free measure of what counts as good enough for generalization. In addition to support from empirical linguistic studies, experimental research with infants (e.g., Shi & Emond 2023) suggests that the Tolerance Principle is a domain-general mechanism and can be applied to social learning and cultural conventionalization, providing a more accurate account of behavioral data than rational decision processes.

 

References

Martínez, H. J. V., Heuser, A. L., Yang, C., & Kodner, J. (2023). Evaluating Neural Language Models as Cognitive Models of Language Acquisition. arXiv:2310.20093. 
Shi, R., & Emond, E. (2023). The threshold of rule productivity in infants. Frontiers in Psychology, 14, 1251124.
Yang, C. (2016). The price of linguistic productivity. MIT Press.
Yang, C., Crain, S., Berwick, R. C.,Chomsky, N., & Bolhuis, J. J. (2017). The growth of language: Universal Grammar, experience, and principles of computation. Neuroscience & Biobehavioral Reviews, 81, 103-119.