Transferência de aprendizagem

A pretrained model provides parameters or representations learned before the target task. Reusing that starting point can be useful when target examples are limited, but it does not guarantee a saving or an accuracy improvement in every setup. In fixed feature extraction, a pretrained base supplies representations while a new prediction head is trained. The base weights remain unchanged by optimizer updates in this stage. Fine-tuning instead permits selected pretrained weights to be updated on the target data. A smaller learning rate is often recommended in the illustrated transfer-learning workflows to limit disruptive parameter changes, especially after adding a randomly initialized head. How many layers to unfreeze and how quickly to update them should be evaluated. Source and target tasks can differ in important ways. Image-network research by Yosinski and colleagues found differences in the transferability of features from different layers, with later features more specialized in their evaluated settings. Task distance affected results, but the finding does not mean every early feature transfers or every distant task fails. Framework behavior introduces additional details. In Keras, a weight's trainable status and the layer's training or inference behavior are generally separate concepts. BatchNormalization is a special case that also stores moving statistics. The cited fine-tuning guide recommends keeping those layers in inference mode when unfreezing the base to avoid disrupting the stored statistics. It also instructs recompiling the Keras model after trainable settings change so the compiled training configuration reflects those choices. These details concern that documented workflow rather than every machine-learning framework.