Fine-tuning
lastmod 2025-10-28created 2025-10-281 edits
In deep learning, fine-tuning is the process of specializing a model designed for one task (the upstream task) to a new task (the downstream task). It is therefore a special case of transfer learning.
Technically, fine-tuning involves applying additional training (e.g., training on new data) to the parameters of a pre-trained artificial neural network. Various variants exist. The additional training may be applied to the entire network, or only to a subset of the hidden layers. In the latter case, the layers that are not fine-tuned are "frozen" (i.e., not changed during backpropagation). A model may also be augmented with "adapters", lightweight modules inserted into the model architecture that fine-tune the embedding space for domain adaptation. These modules contain far fewer parameters than the original model and can be fine-tuned in a parameter-efficient way by adjusting only the adapter weights and freezing the weights of the rest of the model.
For some architectures, such as convolutional neural networks, it is common to keep the early layers closest to the input layer frozen, since they capture low-level features, while later layers identify high-level features that may be more related to the task the model was trained on.
Models pre-trained on large, general corpora are usually fine-tuned by reusing their parameters as a starting point and adding task-specific layers trained from scratch. Fine-tuning the full model is also common and often yields better results, but is more computationally expensive.
Fine-tuning is usually done through supervised learning, but there are also techniques for fine-tuning a model using weak supervision. Fine-tuning can be combined with objectives based on reinforcement learning from human feedback to produce language models such as ChatGPT (a fine-tuned version of a GPT model) and Sparrow.