The child as hacker : building more human-like models of learning

Abstract

Cognitive science faces a radical challenge in explaining the richness of human learning and cognitive development. This thesis proposes that developmental theories can address the challenge by adopting perspectives from computer science. Many of our best models treat learning as analogous to computer programming because symbolic programs provide the most compelling account of sophisticated mental representations. We specifically propose that learning from childhood onward is analogous to a style of programming called hacking--

making code better along many dimensions through an open-ended and internally-motivated set of diverse values and activities. This thesis also develops a first attempt to formalize and assess the child as hacker view through an in-depth empirical study of human and machine concept learning. It introduces list functions as a domain for psychological investigation, demonstrating how they subsume many classic concept learning tasks while opening new avenues for exploring algorithmic thinking over complex structures. It also presents HL, a computational learning model whose representations, objectives, and mechanisms reflect core principles of hacking. Existing work on concept learning shows that learners both prefer simple explanations of data and find them easier to learn than complex ones. The child as hacker, by contrast, suggests that learners use mechanisms that dissociate hypothesis complexity and learning difficulty for certain problem classes.

We thus conduct a large-scale experiment exploring list functions that vary widely in difficulty and algorithmic content to help identify structural sources of learning difficulty. We find that while description length alone predicts learning, predictions are much better when accounting for concepts' semantic features. These include the use of internal arguments, counting knowledge, case-based and recursive reasoning, and visibility--a measure we introduce to modify description length based on the complexity of inferring each symbol in a description. We further show that HL's hacker-like design uses these semantic features to better predict human performance than several alternative models of learning as programming. These results lay groundwork for a new generation of computational models and demonstrate how the child as hacker hypothesis can productively contribute to our understanding of learning.