COMP 2308 · Data Science: Deep Learning

Led by Rumelhartian Parallel Distributed Processing Simulacrum

Feedforward networks, backpropagation, overfitting, MNIST, and an end-to-end business case — deep learning from first principles to production.

1. Module 1

   The Feedforward Network

   Led by Rumelhartian Parallel Distributed Processing Simulacrum

   The question

   Two linear layers (Y = W2(W1X)) are mathematically equivalent to one linear layer. Why does adding a ReLU activation break this equivalence — and why does this make deep networks capable of things shallow networks cannot do?

   Outcome

   The student can describe the feedforward architecture and build a two-hidden-layer network in Keras.

   Sub-units

   1. ○ 1.1 Why Nonlinearity? Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   Backpropagation and Gradient Descent

   Led by Rumelhartian Parallel Distributed Processing Simulacrum

   The question

   Backpropagation computes gradients for a million-weight network in two passes (one forward, one backward). What does the chain rule do — and what does Adam do differently from vanilla gradient descent?

   Outcome

   The student can implement gradient descent in NumPy and explain what Adam adds.

   Sub-units

   1. ○ 2.1 Train in NumPy Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Overfitting, Regularisation, and Validation

   Led by Rumelhartian Parallel Distributed Processing Simulacrum

   The question

   Training accuracy 98%, validation accuracy 75% — the network has memorised the training data. Dropout zeros out neurons randomly. Early stopping halts when validation loss rises. What exactly does each technique do to the loss landscape?

   Outcome

   The student can identify overfitting from training curves and apply dropout and early stopping.

   Sub-units

   1. ○ 3.1 Overfit and Fix Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   MNIST: The Standard Benchmark

   Led by Rumelhartian Parallel Distributed Processing Simulacrum

   The question

   The correct MNIST workflow: normalise, split, train with early stopping, test once at the end. Why does evaluating on the test set multiple times invalidate the result — and what would a 97% test accuracy mean for a deployed digit recogniser?

   Outcome

   The student can implement the complete MNIST pipeline achieving >96% test accuracy.

   Sub-units

   1. ○ 4.1 Build the MNIST Network Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Business Case: Deep Learning End to End

   Led by Rumelhartian Parallel Distributed Processing Simulacrum

   The question

   80% validation accuracy on a churn prediction model. A marketing manager asks: "what should I do with this?" How do you communicate a model's predictions, its limitations, and its actionable implications to a non-technical decision-maker?

   Outcome

   The student can implement and communicate an end-to-end deep learning business case.

   Sub-units

   1. ○ 5.1 Final Report: Business Case Open this unit → ✓ ↻
   Open this module → Print certificate of completion