Not starting from nothing.
WHAT TRANSFER LEARNING IS
Taking a model trained on a large dataset and adapting it to your task.
WHY IT MATTERS
Training from nothing requires enormous data and computation. Adapting requires far less of both.
WHAT IT MAKES POSSIBLE
Useful models from hundreds or thousands of examples rather than millions.
HOW IT WORKS
Keeping the learned general features, and retraining the final layers on your data.
WHAT MODEL REPOSITORIES PROVIDE
Trained models for common tasks, ready to use or adapt.
WHAT TO CHECK BEFORE USING ONE
The licence What data it was trained on Known limitations and biases Its size and inference cost
WHY THE TRAINING DATA MATTERS
A model trained on data unlike your situation performs badly on it.
WHAT THAT MEANS PRACTICALLY
Models trained predominantly on one population may perform worse on others.
WHAT TO DO ABOUT IT
Evaluate on your own data, across the groups you serve.
WHAT TO START WITH ALWAYS
An existing model, before considering training your own.
WHAT TO ESTABLISH
Whether a ready-made API would serve, without any model work.
