Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals

Humans can learn several tasks in succession with minimal mutual interference but perform more poorly when trained on multiple tasks at once. The opposite is true for standard deep neural networks. Here, we propose novel computational constraints for artificial neural networks, inspired by earlier work on gating in the primate prefrontal cortex, that capture the cost of interleaved training and allow the network to learn two tasks in sequence without forgetting. We augment standard stochastic gradient descent with two algorithmic motifs, so-called “sluggish” task units and a Hebbian training step that strengthens connections between task units and hidden units that encode task-relevant information. We found that the “sluggish” units introduce a switch-cost during training, which biases representations under interleaved training towards a joint representation that ignores the contextual cue, while the Hebbian step promotes the formation of a gating scheme from task units to the hidden layer that produces orthogonal representations which are perfectly guarded against interference. Validating the model on previously published human behavioural data revealed that it matches performance of participants who had been trained on blocked or interleaved curricula, and that these performance differences were driven by misestimation of the true category boundary. Author summary Humans can learn multiple tasks over their lifetime with minimal forgetting. In contrast, machine learning architectures based on artificial neural networks fail to learn multiple tasks in sequence and require data of all tasks to be present at once. Previous reports suggest that the opposite is true for humans: We learn better when trained on one task at a time. Here, we sought to identify the basis-set of algorithmic motifs required to mimic human-like continual learning. We propose a novel training method inspired by insights into the function of prefrontal cortex of the human brain. The method consists of task neurons that carry information over successive trials, and an update step that links those to other neurons that encode task-relevant information. Together, these two innovations allow us to model human continual task performance. Analysing how the network represented task information revealed striking similarities between our network and recent reports on task representations in the prefrontal cortex of the mammalian brain. Taken together, our approach describes an effort to bridge insights from machine learning and neuroscience to advance our understanding of the algorithmic basis of continual learning