A team is adapting a pretrained backbone of about 25 million weights to 1,400 photographs across 7 classes. They try three configurations and record the best accuracy each one reaches.
| run | trainable weights | learning rate | training accuracy | validation accuracy |
|---|---|---|---|---|
| A | the new head only | 88% | 85% | |
| B | the new head and the last backbone block | 96% | 90% | |
| C | the new head and the entire backbone | 99.6% | 78% |
Run C is by far the worst on validation.
Which explanation fits the evidence in the table?
Select all that apply.