Machine Learning

Context: FIT1043_MOC · the hacking ∩ maths/stats region of Conway’s diagram · algorithms that learn from data · the Analysis step of the value chain · learns predictive models in two styles

Quick Revision

• 🎯 Objective: build algorithms/techniques that let computers learn from data ➔ patterns/inference, not explicit instructions.
• ⚡ Key Constraint: ML earns its place only when hand-written rules fail — the task is unknown, changing, or too big/expensive for humans.

📝 Core

• Definition ➔ the study of algorithms/statistical models that perform a task without explicit instructions, relying on patterns and inference (Wikipedia); getting computers to learn and improve over time autonomously from data/observations (Emerj); underlying theory in statistics.
• vs expert systems ➔ rule-based expert systems are brittle when the situation keeps changing; ML adapts from data instead.
• When ML is useful ➔
  • No human expertise ➔ e.g. Martian exploration.
  • Tacit expertise ➔ humans can’t state the rules, or they’re incomplete/need tuning ➔ e.g. speech recognition.
  • Auto-adaptation ➔ many solutions must adapt automatically ➔ e.g. user personalisation.
  • Changing conditions ➔ e.g. junk-email filtering.
  • Large data ➔ e.g. discovering astronomical objects.
  • Humans too expensive ➔ e.g. handwritten zipcode recognition.

⚙️ Developing an ML Model

• 1. Choose a measure of success ➔ the metric the model is judged by.
• 2. Set an evaluation protocol ➔ how you’ll test it (e.g. held-out test; see Predictive Models).
• 3. Build a benchmark model ➔ a simple baseline to beat.
• 4. Build a better model + tune hyperparameters ➔ improve on the baseline and adjust its settings.

⚠️ Common Mistakes

• 💡 ML ≠ all of data science ➔ it’s one region of Conway’s diagram (hacking ∩ maths/stats); “ML on big data” is the narrow definition of data science.
• 💡 Rules vs learning ➔ if a simple, stable rule set solves it, a rule-based system may beat ML; reach for ML when rules can’t be written or won’t stay valid.

🧠 Active Recall

Give the agreed definition of machine learning and two situations where it beats a rule-based system.

Answer

• Short answer: ML develops algorithms/techniques that let computers learn from data (statistical foundation); it beats rules when humans can’t express the expertise (speech recognition) or when conditions keep changing (junk email).
• Why: Learn vs encode ➔ rule-based systems fix behaviour up front; ML infers and adapts it from data, handling tacit, shifting, or high-volume problems.

Outline the four steps to develop an ML model.

Answer

• Short answer: (1) choose a measure of success; (2) set an evaluation protocol; (3) build a benchmark/baseline model; (4) build a better model and tune its hyperparameters.
• Why: Baseline-then-improve ➔ a fixed metric + protocol lets you prove each model beats the last, avoiding tuning to noise.