Model Trade-Offs
Interpretability versus explainability
The story: Two ways to understand a machine. With a bicycle, you can look at the chain and gears and see exactly why pedaling moves the wheel. With a modern car engine full of electronics, you can't see inside, so you learn by testing: "when I press the pedal this much, it goes this fast". The bicycle is simple enough to read directly. The engine you can only explain from the outside.
In AI/AWS terms:
| Interpretability (the bicycle) | Explainability (the engine) | |
|---|---|---|
| Meaning | A human can explain the output directly from the model's weights and features | The behavior of any model, even a black box, is explained in human terms |
| How | Inspect the inner mechanics | Model-agnostic methods: partial dependence plots, SHAP dependence plots, surrogate models |
| Example | An economist reads the coefficients of a regression model for inflation | A news site discovers its neural network labels business articles about sports teams as "sports" |
A bicycle is easy to understand but slow. High interpretability usually costs performance. If you need exact transparency, your choice of algorithms is limited. If you need high performance with a general understanding, use explainability.
For the exam: Interpretability = read the model's insides directly, usually at a performance cost. Explainability = explain any model's behavior from the outside.
Safety versus transparency
The story: A bank vault is safe because nobody can see inside. The more windows you add so people can see how it works, the less safe it gets. You're always choosing between the two.
In AI/AWS terms: Model safety is avoiding harm: bias, privacy exposure, and security vulnerabilities. Safety protects information and transparency exposes it, so they pull against each other:
- Accuracy: complex neural networks are more accurate but less interpretable than linear models.
- Privacy: techniques like differential privacy improve safety but make models harder to inspect.
- Safety: filtering outputs hides the original reasoning.
- Security: models trained on air-gapped (fully disconnected) networks are harder to audit externally.
For the exam: Safety and transparency trade off against each other. Differential privacy and air-gapped training raise safety but lower transparency.
Controllability
The story: A recipe is controllable if changing the ingredients changes the dish the way you expect. Add more sugar, it gets sweeter. If adding sugar sometimes makes it saltier, you can't fix a dish that's gone wrong.
In AI/AWS terms: A controllable model is one whose predictions and behavior you can influence by changing the training data. It's more transparent and makes it easier to correct bias.
- Linear models are more controllable than complex neural networks.
- Test it by adding or removing examples and checking that the output changes as expected (the sugar test).
- Improve it with data augmentation and training constraints.
For the exam: Controllability = changing the training data changes the model's behavior predictably.