Fig. 3: Workflow illustrating the fine-tuning process.
From: Generalizable deep learning models for predicting laboratory earthquakes
Initially, the model is pre-trained using the continuous AE signal from the biaxial-loading dataset, with shear force as the target. Post-pretraining, the first five layers of decoder of the deep convolutional model are kept and fixed, and only the Regression Head, constituting 3% of the total model weight, undergoes fine-tuning. This fine-tuning is carried out with a subset of DDS datasets, targeting different parameters like shear stress and time to failure. Finally, the fine-tuned model is applied to predict the remaining DDS datasets.
