A team collects 2,000 photographs of single hands, each held up with the palm facing the camera. Exactly half are left hands and half are right hands. They train two separate copies of the same network on these photographs:
To fight overfitting, both models use the same augmentation pipeline, applied afresh to every training image in every epoch:
| transformation | setting |
|---|---|
| horizontal flip (mirror image) | applied to a random half of the images |
| rotation | a random angle of up to |
| brightness | a random change of up to |
The validation photographs are left untouched. After 30 epochs:
| training accuracy | validation accuracy | |
|---|---|---|
| Model 1 (finger count) | 97% | 94% |
| Model 2 (left or right) | 51% | 50% |
Model 2's two numbers have sat at those values since the second epoch.
What is holding Model 2 at 50%, and what would fix it?
Select all that apply.