Transfer Learning Decision Memo

Use this before using a pretrained model.

Task

• Task:
• Dataset:
• Intended use:
• Why a small model from scratch may not be enough:

Pretrained Model

• Candidate model:
• Source:
• License/terms checked:
• Expected download size:
• Cache behaviour:
• Training data known or partly known:

Decision

• Feature extraction / fine-tuning / do not use:
• Layers frozen:
• Layers trainable:
• Evaluation plan:
• Error-analysis plan:

Risks

• Dataset mismatch:
• Bias or representation risk:
• Privacy or rights risk:
• Overclaiming risk:
• Reproducibility risk:

Boundary

• What this model may show:
• What it must not prove:
• Human review needed before real use:

Mark complete