Supervised learning outperforms unsupervised learning only for categories of intermediate difficulty

Humans learn about the world both when supervision is present and when it is absent. The degree to which existing internal representations align with task demands affects how supervised and unsupervised learning progress. However, mixed results in the literature demand a systematic test of whether supervision is particularly beneficial when tasks align (or misalign) with internal representations. We assessed a large sample (N=662) of participants in an auditory category learning task in which we manipulated supervision and alignment. Our results replicate both that supervised and unsupervised learning are more successful when representations and task demands align, and that supervised learning is generally beneficial. Importantly, our results also indicate that the benefit of supervision is inversely U-shaped, such that it is greatest when tasks are challenging but diminishes when tasks are too easy or too difficult. We explain how our results reconcile the mixed findings in the literature, and speculate how other aspects of learning could further modulate the influence of alignment. Our demonstration that assessment of representation-to-task alignment affects supervised and unsupervised learning to different degrees has implications for instructional design.