Search Papers | Poster Sessions | All Posters
Poster B70 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Easing into learning: How and why a sequential curriculum improves visual learning generalization
Charlotte Volk1,2 (), Christopher Pack1,3, Shahab Bakhtiari2,4; 1McGill University, 2Mila - Quebec AI Institute, 3Montreal Neurological Institute-Hospital, 4Université de Montréal
Generalization in human visual learning (VL) varies across tasks, with 'easy' visual tasks (e.g., large angle orientation discrimination) generalizing better to unseen conditions than 'hard' ones (e.g., small angle orientation discrimination). We used an artificial neural network (ANN) model to explore how training on a sequential curriculum (easy to intermediate to hard) enhances VL generalization. Our findings revealed that the dimensionality of the representational readout subspace, established during the initial training phase, is crucial for generalization. Specifically, 'harder' tasks in later stages can 'piggyback' on the low-dimensional, more generalizable subspace established during the 'easier' initial training phase, leading to a more generalizable outcome.
Keywords: visual learning artificial neural networks plasticity perception