The curriculum effect in visual learning: the role of readout dimensionality

Abstract

Generalization of visual perceptual learning (VPL) to unseen conditions varies across tasks. Previous work suggests that training curriculum may be integral to generalization, yet a theoretical explanation is lacking. We propose an explanatory theory of visual learning generalization and curriculum effects by leveraging an artificial neural network (ANN) model of VPL in comparison with humans. We found that easy-to-hard sequential training improved generalization in both humans and ANNs. However, when easy and hard conditions were interleaved, humans and ANNs showed different behaviours: while ANNs performed worse than with sequential training, humans maintained good performance but with large inter-individual variability. Investigating ANN models trained with different curricula, we demonstrated that models relying on low-dimensional neural populations showed superior generalization. This readout subspace dimensionality was directly determined by curriculum: learners who learned from easy tasks early formed lower-dimensional subspaces and generalized better. Our theory provides a mechanistic framework linking curriculum design to VPL generalization through neural population dimensionality. Author SummaryLearning new skills is fundamental to humans and animals. However, a key challenge in learning is generalization: applying learned skills to new situations not encountered during training. While it is well known that training curriculum affects how well we generalize, the underlying mechanisms remain poorly understood. What makes certain training curricula more effective than others? Why do some learners generalize better than others even with similar training? We developed a computational theory to explain how curriculum design influences generalization in visual learning. Using a combination of human behavioural experiments and artificial neural network modeling, we posit that easy-to-hard training sequences lead learners to focus on fewer, more essential visual features. This creates low-dimensional neural representations that are robust and generalize well to new conditions. Importantly, we found this benefit even extends to learners who spontaneously adopt easy-to-hard strategies on their own, without explicit curriculum design. Our theory provides a mechanistic link between curriculum structure, neural population dynamics, and generalization performance. Moreover, these findings offer practical guidelines for designing effective training programs in education and rehabilitation.