MEPX
Chapter 9 of 10All chapters
  1. 1The artificial neuron
  2. 2Layers and depth
  3. 3How it learns
  4. 4Activation functions
  5. 5Architectures
  6. 6Keeping it general
  7. 7Training in practice
  8. 8Transfer learning
  9. 9What they cannot do
  10. 10Learning it yourself

Chapter 9 of 10

What they cannot do

Honest limitations.

Brittleness

Small deliberate changes to an input can flip a confident prediction, and performance falls sharply on data unlike the training distribution.

  • Adversarial examples remain unsolved rather than merely unfixed.
  • Confidence scores are not reliable measures of correctness.

Interpretability

Behaviour emerges from millions of interacting weights, so there is no rule to read off. Explanation methods give approximations, which matters wherever decisions must be justified.