Why transfer learning saves time and data in TensorFlow - Why Metrics Matter

Metrics & Evaluation - Why transfer learning saves time and data

Which metric matters and WHY

When using transfer learning, accuracy and training time are key metrics. Accuracy shows how well the model learned the new task. Training time shows how much faster transfer learning is compared to training from scratch. Also, data efficiency matters because transfer learning needs fewer new examples to reach good accuracy.

Confusion matrix example

Actual \ Predicted | Positive | Negative
------------------|----------|---------
Positive          |    90    |   10    
Negative          |    15    |   85    

This confusion matrix shows a model trained with transfer learning. It correctly predicts 90 positives and 85 negatives out of 200 samples.

Precision vs Recall tradeoff with examples

Transfer learning helps balance precision and recall quickly. For example, if you want to detect cats in photos, transfer learning can quickly improve recall (finding most cats) without losing precision (not mistaking dogs for cats). This saves time and data compared to training a new model from scratch.

Good vs Bad metric values for transfer learning

• Good: Accuracy above 85% with less than 10% of original training data and training time reduced by 50% or more.
• Bad: Accuracy below 70% or needing almost the full original dataset and long training times, meaning transfer learning did not save time or data.

Common pitfalls in metrics

• Assuming transfer learning always improves accuracy without checking if the source and target tasks are related.
• Ignoring overfitting signs when fine-tuning on small data.
• Confusing faster training time with better model quality.
• Data leakage from source task data into target task evaluation.

Self-check question

Your transfer learning model reaches 90% accuracy using only 20% of the original data and trains in half the time. Is this good?
Answer: Yes, this shows transfer learning saved time and data while keeping good accuracy, which is the goal.

Key Result

Transfer learning saves time and data by achieving good accuracy faster with less new data.