Editing Transfer Learning (section)

== <span style="color: #FFFFFF;">Evaluating</span> ==
Expert transfer learning practitioners evaluate along a specific set of dimensions:

'''Transfer gain''': Compare fine-tuned model vs. training-from-scratch baseline on the same target data. The transfer gain is the performance improvement attributed to the pre-trained initialization. If fine-tuning doesn't beat scratch after enough epochs, transfer learning may not be helping.

'''Few-shot evaluation curves''': Plot performance as a function of available labeled data. Transfer learning should show much better performance at low data regimes, converging toward the from-scratch baseline as data increases.

'''Feature analysis''': Using linear probing — train only a linear classifier on frozen pre-trained features — to measure how much task-relevant information is already encoded in the pre-trained representations before any fine-tuning.

'''Domain proximity metrics''': Measure the distribution distance (FID, A-distance, or H-divergence) between source and target domains. Higher distance predicts lower transfer benefit and may suggest negative transfer risk.

Expert practitioners also perform '''layer-wise learning rate sweeps''' to identify which layers benefit most from fine-tuning, rather than applying a blanket strategy.
</div>

<div style="background-color: #2F4F4F; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">