Failure to learn during roving, analysing the unsupervised bias hypothesis

We examine the unsupervised bias hypothesis [11] as an explanation for failure to learn two bisection tasks, when task sequencing is randomly alternating (roving). This hypothesis is based on the idea that a covariance based synaptic plasticity rule, which is modulated by a reward signal, can be biased when reward is averaged across multiple tasks of differing difficulties. We find that, in our hands, the hypothesis in its original form can never explain roving. This drives us to develop an extended mathematical analysis, which demonstrates not one but two forms of unsupervised bias. One form interacts with overlapping task representations and the other does not. We find that overlapping task representations are much more susceptible to unsupervised biases than non-overlapping representations. Biases from non-overlapping representations are more likely to stabilise learning. But this in turn is incompatible with the experimental understanding of perceptual learning and task representation, in bisection tasks. Finally, we turn to alternative network encodings and find that they also are unlikely to explain failure to learn during task roving as a result of unsupervised biases. As a solution, we present a single critic hypothesis, which is consistent with recent literature and could explain roving by a, much simpler, certainty normalised reward signalling mechanism.