A frozen layer refers to a specific layer within an intelligence artificielle (AI) model, particularly in apprentissage profond frameworks, that has been set to remain unchanged during the training process. This means that the weights and biases of this layer are not updated when the model is trained on nouvelles données. Freezing layers is a common practice in l'apprentissage par transfert, where a model pre-trained on a large dataset is adapted to a smaller, task-specific dataset.
By freezing certain layers, developers can preserve the learned features that those layers have captured, which may be beneficial for similar tasks. For example, in a réseau de neurones convolutionnels (CNN) used for image recognition, earlier layers might capture low-level features such as edges and textures, while deeper layers capture more complex patterns. Freezing the earlier layers allows the model to retain these foundational features while only training the later layers to adapt to the new dataset.
Freezing layers can also help in speeding up the training process, as fewer parameters need to be adjusted. Additionally, it can combat overfitting, especially when the new dataset is small compared to the original dataset used for pre-training. However, careful consideration is required to determine which layers to freeze and which to keep trainable, as this decision significantly impacts performance du modèle.