問題文
A team adapts a large pretrained image-and-text model to a narrow inspection task with limited compute. They cannot afford to update all the weights. Which approach keeps the adaptation cheap while still changing behavior?
選択肢
- Update only the final classification layer and leave every other layer frozen, because the earlier layers encode generic features and the final layer alone carries all the task-specific behavior that the adaptation needs to change for this inspection task.
- Train on a smaller subset of the labeled images, so each epoch touches fewer examples and the full update becomes affordable.
- Reduce the image resolution so the full update fits the memory budget, which lowers the cost of every layer at once.
- Train a small number of added parameters while keeping the pretrained weights frozen.