Fig 6. Modelling the benefit of blocked over interleaved training observed in data from human participants.

(A) Test phase accuracy. Humans perform better under blocked compared to interleaved training, while the baseline model performs better under interleaved training, due to catastrophic forgetting. Our Hebbian network with sluggish task units performs better under blocked training. (B) Congruency effect. Just like human participants, under interleaved training, our model performs worse on incongruent compared to congruent trials. (C) Sigmoidal fits of choices made by human participants, the baseline network, and our network. Our network recreates the intrusion of irrelevant dimensions observed under interleaved, but less so under blocked training. (D) Fits of the factorised and linear model to human and network choices. In contrast to the baseline model, our model recreates patterns observed in humans, where the factorised model fits better under blocked than interleaved training.