Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals
Fig 2
Task design, network architecture and results of blocked versus interleaved learning.
(A) Task design. Stimuli were two-dimensional Gaussian functions (“blobs”) for which we systematically varied the location of its peak along the x- and y- dimensions in five discrete steps. Each subpanel visualises the Gaussian blob input image at that location in the underlying 2D stimulus space. Only one of the two feature dimensions was relevant per task, so that the reward (y-label) depended on the x-position in the first task (orange) and y-position in the second task (blue). (B) The network was a simple feed-forward MLP with a single hidden layer with ReLU non-linearities and received the flattened images of Gaussian blobs together with a one-hot encoded task signal as inputs. (C) The network was trained either in a fully interleaved curriculum in which trials from both contexts were randomly interspersed, or in a blocked curriculum in which it was first trained on one task, and then on the other. (D) Under interleaved training, the network quickly reached 100% training accuracy on both tasks. In contrast, under blocked training, learning the second task came at the cost of forgetting how to perform the first task. (E) Plotting the choices of the trained network in two dimensions revealed that under interleaved training, choices were aligned with the ground truth category boundaries (shown in (A)), whereas under blocked training, the network treated the first task as if it was the second. (F) Projections of the hidden layer activity into three dimensions via multi-dimensional scaling (MDS) shows orthogonal representations under interleaved training, where irrelevant information was suppressed, and parallel representations under blocked training, where the first task is encoded in the same way as the second task. (G) Under interleaved training, a significant proportion of hidden units were exclusively selective to the relevant dimension in one task (but not the other), whereas no such task-selectivity was observed under blocked training. (H) Evolution of correlation between task weights for both tasks during training. Interleaved—but not blocked—training promoted learning of anti-correlated task weights.