A pretrained image model already recognizes general visual features, but the company has a small labeled dataset for its own product categories. Which approach should be evaluated first?

1. Report only performance on the fine-tuning examples.
2. Transfer learning from a suitable pretrained model with a held-out task evaluation.
3. Use the original model labels unchanged even though the business categories differ.
4. Train a very large model from random initialization without comparing a pretrained baseline.