Short answers, pulled from the story.
Fine-tuning begins when a neural network trained for one task meets new data. The original model known as the upstream task must now perform a different downstream task. This process reuses knowledge from the initial training objective to solve specific problems.
Designers create a low-rank matrix and add it to the original matrix to produce a fine-tuned model when combined with a base model. A language model containing billions of parameters might be LoRA fine-tuned using only several millions of parameters. Support for LoRA was integrated into the diffusers library from Hugging Face.
Convolutional neural networks often keep earlier layers frozen because they capture lower-level features. These early layers sit closest to the input layer and handle basic pattern recognition while later layers discern high-level features that relate more directly to the specific task at hand.
Fine-tuning can degrade a model's robustness to distribution shifts in real-world scenarios. One mitigation strategy involves linearly interpolating a fine-tuned model's weights with the original model's weights. This process greatly increases out-of-distribution performance while largely retaining the in-distribution performance of the fine-tuned model.
As of the 19th of June 2023, language model fine-tuning APIs are offered by OpenAI for a subset of their models. Microsoft Azure's Azure OpenAI Service provides similar capabilities for selected models. Google Cloud Platform also supports fine-tuning for some of their PaLM models.