Memorization Indicators Spot Overfitting in Low-Sample sEMG Calibration

Deep learning models for surface electromyography (sEMG) often require subject-specific calibration to achieve optimal performance, as large, diverse generic datasets are scarce. However, the practical constraint of collecting limited data during calibration significantly increases the risk of overfitting, potentially degrading model performance. Traditional methods for detecting overfitting, such as validation performance monitoring and early stopping, are challenging to apply in these low-sample scenarios due to the lack of sufficient held-out data. This study explores a novel class of memorization indicators that rely solely on the activation statistics of rectified linear units (ReLU) within deep neural networks. These indicators can be computed directly from training data, eliminating the need for a separate validation set. Through a transfer learning experiment on an sEMG benchmark dataset, where a pre-trained convolutional neural network is fine-tuned on individual users with minimal repetitions, researchers observed a clear correlation. Decreases in test accuracy during fine-tuning were accompanied by distinct changes in activation rates, suggesting that these activation-based indicators are a viable tool for early detection of unsuccessful learning in sEMG calibration with limited data.