You download a pretrained image model and adapt it to a new job: sorting photographs of machine parts into eight categories.
What the downloaded model contains:
| quantity | value |
|---|---|
| parameters in the whole downloaded model | |
| length of the feature vector the backbone produces | numbers |
| classes the downloaded head predicts |
The head is a single fully connected layer: one weight for every (input number, output class) pair, plus one bias per class. The backbone is everything underneath it.
You throw the 1000-class head away, bolt on a new head of exactly the same design sized for your eight classes, keep every backbone weight exactly as it was downloaded, and train only the new head.
How many parameters does that training update?
Select all that apply.