Improving the repeatability of two-rate model parameter estimations by using autoencoder networks.

The adaptive changes elicited in visuomotor adaptation experiments are usually well explained at group level by two-rate models (Smith et al., 2006), but parameters fitted to individuals show considerable variance. Data cleaning can mitigate this problem, but the assumption of smoothness can be problematic due to fast adaptive changes with discontinuous derivatives. In this paper, we collected time-series data from an experimental paradigm involving repeated training and investigated the effect of various cleaning methods, including an autoencoder network (AE), on the parameter estimation. We compared changes in the fitted parameters across different methods and across training repetitions. The results suggest that AE performed best overall, without introducing an underestimation bias on b like moving average or piecewise polynomials, and that it reduced the within-subject variance overall and especially that of the fast retention rate a by >50%.