Training curricula and structured representations in human and machine learning

Thesis

Humans display a remarkable ability to learn from previous experience. Far from being passively received, however, the training experiences we encounter in modern life are often chosen by ourselves or other people. Existing work shows that some training schedules (“curricula”) work better than others in specific situations, but our understanding of when and why these curricula work is limited, making it difficult to predict curriculum success in novel settings. Compounding this issue, the curricula that subjectively feel most effective are not always optimal for promoting long-term retention or generalization. The main aim of this thesis is to contribute towards the development of computationally-informed theories of curriculum efficacy and mechanism. For this purpose, we make use of behavioral experiments, neural recordings and supervised neural network models.

A key focus of the thesis is on curricula that “start small” by constraining the information that is initially available to the learner. This premise is evaluated in three distinct learning domains. For continuous discriminations such as parametric category structures, initially limiting training to prototypical examples helps. We show that this is at least in part because human internal representations are inherently noisy, and prototypicality acts as a buffer against noise-induced distortions. In tasks that require integrating information from multiple cues, the cue properties emphasized by the training curriculum predict the type of decision rule that human learners use. In a compositional rule learning setting, human learners, but not artificial neural networks, benefit from training on individual rules one at a time. A second aspect of study relates to the fact that in most settings, it is not memorization of training examples, but generalization and transfer to novel problems that are most crucial to adaptive behavior. In this regard, we revive an old idea in cognitive science, that human thought and action are fundamentally compositional. Evaluating models with an explicitly modular structure on the same compositional rule learning task as humans, we establish several correspondences. These include sensitivity to training curriculum, and generalization patterns exhibited by humans, but not by traditional artificial neural networks. Taken together, it is hoped that the work presented in this thesis provides a small step forward in understanding the training schedules that facilitate successful learning.

Authors: Dekker, RB

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: compositionality; learning; curriculum learning; neural network; representation learning; computational modeling