Demystifying Machine Learning

Summary:

• You can consider the scores/errors on the training or test sets.
• High training error:
  • Possibly underfitting (high bias)
  • Bias can be seen as the training error
• Low training error, high test error:
  • Possibly overfitting (high variance)
  • Variance can be seen as the difference between the training error and the test error.
• Aim for low training and test error (the ‘sweet spot’ - low bias and low variance).