Traditional Programming vs ML

Complexity in Engineering Systems

• Automobile: ~30,000 parts; Airplane: ~3,000,000 parts
• MS Office: ~ 40,000,000 LOCs; Debian: ~ 400,000,000 LOCs

Q. How do we build such complex systems?

Managing Complexity in Software

• Abstraction: Hide details & focus on high-level behaviors
• Reuse: Package into reusable libraries & APIs with well-defined contracts
• Composition: Build large components out of smaller ones

class Algorithms {
    /**
     * Finds the shortest distance between to vertices.
     * This method is only supported for connected vertices.
     */
    int shortestDistance(Graph g, Vertice v1, v2) {…}
}

Contracts in ML?

Q. Is this the same kind of contract as in software?

(Lack of) Modularity in ML

• Often no clear specification of "correct" behavior
  • Optimizing metrics instead of providing guarantees
• Model behavior strongly dependent on training & test sets
  • What happens if distribution changes?
• Poorly understood interactions between models
  • Ideally, develop models separately & compose together
  • In general, must train & tune together

Concept Drifts

• ML estimates "f(x) = y"
  • What if the relationship between "x" & "y" changes over time?
  • What if "f" does not capture certain relationships?
• Q. Examples?
• In general, impossible to predict
  • Continuously monitor and update model

Feedback loops

• Every system is deployed as part of an environment
• Output influences the environment
  • In turn, affects input back to the system
  • Over time, may lead to undesirable (and difficult to reverse) outcome
  • Higher risks if initial data set & model is biased

Example: Crime Prediction

• Use past data to predict crime rates
• Police increases the frequency of patrol in area X
• More arrested made in area X
• New crime data fed back to the model
• Repeat

Discussion: Product Recommendations

• Specification/metrics?
• Concept drift?
• Feedback loop?

What is Measurement?

• Measurement is the empirical, objective assignment of numbers, according to a rule derived from a model or theory, to attributes of objects or events with the intent of describing them. – Craner, Bond, “Software Engineering Metrics: What Do They Measure and How Do We Know?"

• A quantitatively expressed reduction of uncertainty based on one or more observations. – Hubbard, “How to Measure Anything …"

Measurement for Decision Making

• Which project to fund?
• Need more system testing?
• Need more training?
• Fast enough? Secure enough?
• Code quality sufficient?
• Which features to focus on?
• Developer bonus?
• Time and cost estimation? Predictions reliable?

Measurement Scales

• Scale: The type of data being measured.
  • Dictates what sorts of analysis/arithmetic is legitimate or meaningful.
• Nominal: Categories
  • e.g., biological species, film genre, nationality
• Ordinal: Order, but no meaningful magnitude
  • Difference between two values is not meaningful
  • Even if numbers are used, they do not represent magnitude!
  • e.g., weather severity, complexity classes in algorithms
• Interval: Order, magnitude, but no definition of zero.
  • 0 is an arbitrary point; does not represent absence of quantity
  • Ratio between values are not meaningful
  • e.g., temperature (C or F)
• Ratio: Order, magnitude, and zero.
  • e.g., mass, length, temperature (Kelvin)

Measurement Scales

Measurements in AI-Enabled Systems

• Organizational objectives
  • e.g., revenues, growth, lives saved, societal benefits
  • Often not directly measurable from system output; slow indicators
• Leading indicators
  • Customer sentiment: Do they like the product?
  • Customer engagement: How often do they use the product?
  • But can be misleading (more daily active users => higher profits?)
• User outcomes
  • Does the system achieve what it promises to users?
• Model Properties
  • Accuracy of predictions, error rates
  • Performance (e.g., prediction time)
  • Cost: Training time, amount of data required

Exercise: Metrics in Product Recommender

• Organization objectives?
• Leading indicators?
• User outcomes?
• Model properties?
• What are their scales?

The streetlight effect

• A type of observational bias
• People tend to look for something where it’s easiest to do so.

Risks with Measurements

• Bad statistics: A basic misunderstanding of measurement theory and what is being measured.
• Bad decisions: The incorrect use of measurement data, leading to unintended side effects.
• Bad incentives: Disregard for the human factors, or how the cultural change of taking measurements will affect people.

Risks of Metrics as Incentives

• Metrics-driven incentives can:
  • Extinguish intrinsic motivation
  • Diminish performance
  • Encourage cheating, shortcuts, and unethical behavior
  • Become addictive
  • Foster short-term thinking
• Often, different stakholders have different incentives!

Another Case Study: University Rankings

• Originally: Opinion-based polls, but schools complained
• Data-driven model: Rank colleges in terms of "educational excellence"
• Input: SAT scores, student-teacher ratios, acceptance rates, retention rates, alumni donations, etc.,

Discussion: University Rankings

• What is (not) being measured? Any streetlight effect?
• Is the measured data being used correctly?
• Are incentives for using these data good? Can they be misused?

Examples: Risks of Metrics as Incentives

• Example 1
  • Schools optimize metrics for higher ranking (add new classrooms, nicer facilities)
  • Tuition increases, but is not part of the model!
  • Higher ranked schools become more expensive
  • Advantage to students from wealthy families
• Example 2
  • A university founded in early 2010's
  • Math department ranked by US News as top 10 worldwide
  • Top international faculty paid $$ as a visitor; asked to add affiliation
  • Increase in publication citations => skyrocket ranking!

Measurement Validity

• Construct: Are we measuring what we intended to measure?
  • Does the abstract concept match the specific scale/measurement used?
  • e.g., IQ: What is it actually measuring?
  • Other examples: Pain, language proficiency, personality...
• Predictive: The extent to which the measurement can be used to explain some other characteristic of the entity being measured
  • e.g., Higher SAT scores => higher academic excellence?
• External validity: Concerns the generalization of the findings to contexts and environments, other than the one studied
  • e.g., Drug effectiveness on test group: Does it hold over the general public?

Correlation vs Causation

Correlation vs Causation

• In general, ML learns correlation, not causation
  • (exception: Bayesian networks, certain symbolic AI methods)
• To establish causality:
  • Develop a theory ("X causes Y") based on domain knowledge & independent data
  • Identify relevant variables
  • Design a controlled experiment & show correlation
  • Demonstrate ability to predict new cases

Confounding Variables

• If you want to show correlation between X and Y:
  • Identify potential confounding variables
  • Control for those variables during measurement
• Examples
  • Drink coffee => Pancreatic cancer?
  • Degree from high-ranked schools => Higher-paying jobs?

Successful Measurement Program

• Set solid measurement objectives and plans.
• Make measurement part of the process.
• Gain a thorough understanding of measurement.
• Focus on cultural issues.
• Create a safe environment to collect and report true data.
• Cultivate a predisposition to change.
• Develop a complementary suite of measures.

Recitation This Week

Apache JMeter: Performance Measuring Tool