Stochastic Process Model derived indicators of overfitting for deep architectures: Applicability to small sample recalibration of sEMG decoders

Our recent work presents a stochastic process model of the activations within an ANN and shows a promising indicator to distinguish memorizing from generalizing ANNs. The average λ, or mean firing rate (MFR), of a hidden layer, shows stable differences between memorizing and generalizing networks, comparatively independent of the underlying data used for evaluation.
We first show the performance of this indicator during training on benchmark computer vision datasets such as MNIST and CIFAR-10. In a second step, we extend the work to the real-life use case of calibrating a pre-trained model to a new user. We focus on decoding surface electromyographic (sEMG) signals, which are highly variable within and between users, and therefore necessitate frequent user calibration. Especially in situations when user calibration has to only rely on a small number of samples, degradation in performance overtime due to memorization and overfitting is a not unlikely outcome. In those cases, traditional regularization methods that function by observing the performance on a validation set, such as early stopping, don’t necessarily work, because they are evaluated on data from the same subject and set of movements, which features are being memorized. Our new indicators of memorization could help as stable indicators for model performance and give live insights during model calibration when more samples from the new users would be necessary. We evaluate the usefulness of the MFR-indicator for identifying the moment a pre-trained sEMG decoder starts to memorize given inputs