Your next step

Learning that works on new examples

Judge generalisation with separate evidence.

What you’ll learn

Training examples are not the final exam
Identify why evaluation needs unseen examples.
About 4 minutes · Open activity
Validation helps choose; testing checks later
Assign training, validation and test sets distinct roles.
About 4 minutes · Open activity
Too little learned structure
Recognize underfitting from supplied error patterns.
About 4 minutes · Open activity
Remembering quirks instead of patterns
Recognize overfitting from divergent training and validation results.
About 4 minutes · Open activity
Discourage brittle fits
Explain regularization as a way to constrain learning.
About 5 minutes · Open activity
Know when to stop training
Select a checkpoint using supplied validation evidence.
About 4 minutes · Open activity
When the answer leaks into the test
Identify data leakage from a duplicated or unavailable clue.
About 4 minutes · Open activity
The world can change after testing
Recognize distribution shift.
About 4 minutes · Open activity
Choose the trustworthy training result
Inspect training and validation curves, catch an overlapping test example, and name a fresh evaluation for a new setting.
About 6 minutes · Open activity