Decomposed Inductive Procedure Learning: Learning Academic Tasks with Human-Like Data Efficiency

TL;DR

The paper addresses the gap between human-like data efficiency and data-hungry neural methods by proposing Decomposed Inductive Procedure Learning (DIPL), which combines how-, where-, and when-learning into modular, skill-specific induction. Through ablation experiments in two ITS domains, the study shows that single-mechanism RL is data-inefficient, while a two-mechanism setup dramatically improves learning speed, and a full three-mechanism DIPL achieves near-human performance within a small number of problems. Relative featurization and the modular cooperation of mechanisms further enhance efficiency, with DIPL approaching human learning rates and reaching very low error after about 130 problems. The results suggest that integrating multiple specialized learning mechanisms can bridge the gap between human and machine learning, offering a principled path for more data-efficient AI. This work highlights the value of theory-driven cognitive modeling to inform and improve data-efficient learning in ML systems.

Abstract

Human learning relies on specialization -- distinct cognitive mechanisms working together to enable rapid learning. In contrast, most modern neural networks rely on a single mechanism: gradient descent over an objective function. This raises the question: might human learners' relatively rapid learning from just tens of examples instead of tens of thousands in data-driven deep learning arise from our ability to use multiple specialized mechanisms of learning in combination? We investigate this question through an ablation analysis of inductive human learning simulations in online tutoring environments. Comparing reinforcement learning to a more data-efficient 3-mechanism symbolic rule induction approach, we find that decomposing learning into multiple distinct mechanisms significantly improves data efficiency, bringing it in line with human learning. Furthermore, we show that this decomposition has a greater impact on efficiency than the distinction between symbolic and subsymbolic learning alone. Efforts to align data-driven machine learning with human learning often overlook the stark difference in learning efficiency. Our findings suggest that integrating multiple specialized learning mechanisms may be key to bridging this gap.

Figures (6)

• Figure 1: Decomposition from 1-mechanism learning, like RL, that maps states to actions to DIPL’s 3-mechanism learning. A 2-mechanism system bridges the difference and uses how-learning but combines where- and when-learning.
• Figure 2: How-learning explanations for worked examples in fractions and multi-column addition. Of the several explanations, some may be correct (green) or incorrect (red).
• Figure 3: An example of a where-part pattern generalized to act across columns in multi-column addition.
• Figure 4: Positive and negative examples of an AddNum skill in fraction addition, and an Add2 skill in Multi-Column addition. Partial when-parts that accept the positive examples but reject the negative example are shown expressed with relational features generated by relative featurization.
• Figure 5: Fraction arithmetic tutoring system for teaching multiplying and adding fractions. Learner insert an 'x' to the 'check_convert' field if the fraction must first be converted.