A machine learning system is designed to classify medical images. First, a large, pre-existing model, which has a fixed set of internal parameters, processes an image to extract key features. Second, a newly created, smaller component takes these features as input and makes a final prediction (e.g., 'benign' or 'malignant'). Only the parameters of this new, smaller component are adjusted during the training process. Given the general form $Predict_{\tilde{\omega}}(FeatureExtractor_{\tilde{\theta