The limitations of automatically generated curricula for continual learning

In many applications, artificial neural networks are best trained for a task by following a curriculum, in which simpler concepts are learned before more complex ones. This curriculum can be hand-crafted by the engineer or optimised like other hyperparameters, by evaluating many curricula. However, this is computationally intensive and the hyperparameters are unlikely to generalise to new datasets. An attractive alternative, demonstrated in influential prior works, is that the network could choose its own curriculum by monitoring its learning. This would be particularly beneficial for continual learning, in which the network must learn from an environment that is changing over time, relevant both to practical applications and in the modelling of human development. In this paper we test the generality of this approach using a proof-of-principle model, training a network on two sequential tasks under static and continual conditions, and investigating both the benefits of a curriculum and the handicap induced by continuous learning. Additionally, we test a variety of prior task-switching metrics, and find that in some cases even in this simple scenario the a network is often unable to choose the optimal curriculum, as the benefits are sometimes only apparent with hindsight, at the end of training. We discuss the implications of the results for network engineering and models of human development.