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

Area of Science:

• Cognitive psychology
• Neuroscience
• Machine learning

Background:

• Visuomotor adaptation is typically modeled using two-rate models at the group level.
• Individual parameter estimation in these models suffers from significant variance.
• Traditional data cleaning methods may struggle with the non-smooth nature of rapid adaptive changes.

Purpose of the Study:

• To investigate the impact of various data cleaning methods on parameter estimation in visuomotor adaptation.
• To evaluate the performance of an autoencoder network (AE) against conventional cleaning techniques.
• To assess the reduction in within-subject variance, particularly for the fast retention rate (a_f).

Main Methods:

• Collected time-series data from a visuomotor adaptation experimental paradigm with repeated training.
• Applied several data cleaning methods, including moving average, piecewise polynomials, and an autoencoder network (AE).
• Compared parameter estimation and within-subject variance across different cleaning methods and training repetitions.

Main Results:

• The autoencoder network (AE) demonstrated superior performance overall compared to other methods.
• AE did not introduce an underestimation bias on the parameter b_f, unlike moving average and piecewise polynomial methods.
• AE significantly reduced within-subject variance, with a reduction of over 50% for the fast retention rate (a_f).

Conclusions:

• Autoencoder networks offer a robust solution for cleaning visuomotor adaptation data, enhancing parameter estimation accuracy.
• AE methods mitigate the issue of high variance in individual parameters, improving model reliability.
• The application of AE in analyzing adaptation data can lead to more precise understanding of learning dynamics, especially rapid changes.