Revision

Back to ML System Design

Baseline

• Random baseline:
  • Predict at random:
    • Uniform,
    • Following label distribution,
• Zero rule baseline:
  • Always predict the most common class,
• Simple heuristics:
  • E.g.: classify tweets based on whether they contain links to unreliable sources,
• Human baseline:
  • What’s human-level performance?
• Existing solutions.

Evaluation methods

1. Perturbation Test
2. Invariance Tests
3. Directional Expectation Tests
4. Model Calibration
5. Confidence Measurement
6. Slice-based Evaluation

Perturbation Tests

Problem: users input might contain noise, making it different from test data

Examples:

• Speech recognition: background noise
• Object detection: different lighting
• Text inputs: typos, intentional misspelling (e.g. looooooooong)

Model does well on test set, but fails in production.

• Motivation: users input might contain noise, making it different from test data
• Idea: randomly add small noise to test data to see how much outputs change

The more sensitive the model is to noise:

• The harder it is to maintain
• The more vulnerable the model is to adversarial attacks

Solutions

If small changes cause model’s performance to fluctuate, you might want to make model more robust:

• Add noise to training data,
• Add more training data,
• Choose another model.

Invariance Tests

Motivation: some input changes shouldn’t lead to changes in outputs:

• Changing race/gender info shouldn’t change predicted approval outcome,
• Changing name shouldn’t affect resume screening results.

Idea:

• Keep certain features the same, but randomly change values of sensitive features.

Directional Expectation Tests

Motivation: some changes to inputs should cause predictable changes in outputs

Example, when predicting housing prices:

• Increasing lot size shouldn’t decrease the predicted price,
• Decreasing square footage shouldn’t increase the predicted price.

Model Calibration

If you predict team A wins in A vs. B match with 60% probability:

• In 100 A vs. B match, A should win 60% of the time!

Among all samples predicted POSITIVE with propa 80%, 80% of them should be POSITIVE!

Confidence Measurement

• Usefulness threshold for each individual prediction,

• Uncertain predictions can cause annoyance & catastrophic consequences.

• How to measure the confidence level of each prediction?
• What to do with predictions below the confidence threshold?
  • Skip,
  • Ask for more information,
  • Loop in humans.

Slice-based Evaluation

• Classes
  • Might perform worse on minority classes
• Subgroups
  • Gender
  • Location
  • Time of using the app
  • etc.
• Evaluate your model on different slices
• Check for consistency over time

Pros

• Improve model’s performance both overall and on critical data
• Help avoid biases
• Even when you don’t think slices matter, slicing can:

Resources

See:

• CS329, lecture 7.