IDC Revision

HW-02

Original PDF ↗
  1. 1

    The 10-Minute ML Explainer. Work with a partner to explain one specific Machine Learning term, showing that you understand the theory and can explain it with a real-world example. Prepare a 10-minute presentation (5 minutes per person), divided into two parts (or focus on one):

    1. The Theory — What is the concept and why is it a problem?
    2. The Example — Show a real scenario where this happens and how to fix it.

    Format: live presentation. Choose a topic that interests you both (see the suggested list in the body below).
Assignment type

Open-ended presentation, not a worked problem set. There is no marked solution — the deliverable is a live 10-minute explainer.

How you'll be graded

  • Simplicity — could another student understand this easily?
  • Accuracy — is the technical explanation correct?
  • Visuals — clear drawings or charts instead of long sentences?
  • Teamwork — did both partners speak and connect their parts?

Check for understanding (before you start)

  1. The "So What?" test — if this ML problem happens, what is the actual damage to the project?
  2. The "Analogy" test — can you explain the concept to a non-engineer with a simple analogy? (e.g. overfitting is like a student memorising a practice test but failing the real exam.)

Suggested topics

1. Training dynamics & model behaviour ("the fit")

  1. Overfitting — memorising the noise vs. learning the pattern
  2. Underfitting — why a model might be too simple for the data
  3. The bias-variance tradeoff — balancing flexibility with stability
  4. Early stopping — finding the right moment to stop training
  5. Regularization (L1/L2) — a penalty that stops the model over-complicating
  6. Data leakage — when the model "cheats" by seeing test answers early
  7. Hyperparameter tuning — the difference between learning and configuring a model

2. Niche metrics (beyond accuracy)

  1. Precision vs. recall — the cost of a false alarm vs. a missed event
  2. F1-score — the best metric when data is unbalanced
  3. Confusion matrix — a visual map of where the model gets confused
  4. Log loss — how confident (and how wrong) a prediction is
  5. Inference latency — when a slow model is a useless model
  6. False positives vs. false negatives — consequences in medicine vs. security

3. MLOps & real-world operations

  1. Model drift — the world changes but your model doesn't
  2. Data drift — incoming data looks different from training data
  3. Training-serving skew — model behaves differently in the lab vs. production
  4. Model monitoring — an alarm system for AI
  5. A/B testing for AI — comparing two models on real users safely
  6. Feedback loops — a model's own mistakes pollute its future training data
  7. Feature stores — a consistent data library for all your models

4. Data challenges & handling

  1. Imbalanced data — the interesting event is 0.1% of the data
  2. Data augmentation — artificial variety to make the model more robust
  3. Outlier detection — which points are "special" and which are "errors"
  4. Label noise — training when human-labelled data is messy
  5. Cross-validation — proving success isn't just a lucky accident

5. Advanced niche concepts (simplified)

  1. Transfer learning — take a pre-trained brain and give it a new job
  2. Explainable AI (XAI) — asking the black box to show its work
  3. Quantization — shrinking a model to fit on a small device
  4. Retrieval-augmented generation (RAG) — a knowledge base to prevent hallucinations
  5. Adversarial attacks — small, invisible changes that trick a model