COMP 2207 · Machine Learning: Deep Learning

Led by Rosenblattian Neural Network Simulacrum

From Rosenblatt's Perceptron to deep CNNs — backpropagation, overfitting, and convolutional networks. Based on the writings of Frank Rosenblatt.

1. Module 1

   The Perceptron and the ANN Architecture

   Led by Rosenblattian Neural Network Simulacrum

   The question

   A single Perceptron cannot learn XOR. Add one hidden layer and it can. Why — and what does the universal approximation theorem actually say about what deep networks can (and cannot) represent?

   Outcome

   The student can describe the Perceptron, explain why hidden layers overcome XOR, and implement three activation functions.

   Sub-units

   1. ○ 1.1 From Perceptron to Deep Network Open this unit → ✓ ↻
   2. ○ 1.2 Activation Functions Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   Training: Backpropagation and Gradient Descent

   Led by Rosenblattian Neural Network Simulacrum

   The question

   Forward pass, compute loss, backward pass, update weights. The chain rule distributes the gradient. Adam adjusts learning rates per parameter. What do training and validation loss curves tell you about whether learning is working?

   Outcome

   The student can implement a Keras training loop and interpret loss curves.

   Sub-units

   1. ○ 2.1 Train an ANN on the Churn Dataset Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Regularisation and Evaluation

   Led by Rosenblattian Neural Network Simulacrum

   The question

   Dropout randomly zeroes neurons during training to prevent co-adaptation. Early stopping halts training when validation loss stops improving. The confusion matrix on the test set tells you whether it generalised. What does overfitting look like — and what does regularisation do to the loss curves?

   Outcome

   The student can apply dropout and early stopping and interpret confusion matrix results.

   Sub-units

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

   Convolutional Neural Networks

   Led by Rosenblattian Neural Network Simulacrum

   The question

   A 224×224 image has 150,528 pixels. Connecting every pixel to every neuron requires 150 million parameters per layer. Convolutional filters apply the same learned weights at every spatial position. Why does weight sharing work — and what does a trained filter actually detect?

   Outcome

   The student can build a CNN, explain weight sharing, and train on binary image classification.

   Sub-units

   1. ○ 4.1 Build a CNN for Image Classification Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Deep Learning in Practice

   Led by Rosenblattian Neural Network Simulacrum

   The question

   10,000 customer records with 20 features: deep learning or gradient-boosted tree? 500,000 product images: deep learning or not? The data requirement, computational cost, and interpretability requirements all determine the answer.

   Outcome

   The student can identify when deep learning outperforms traditional ML and justify the selection.

   Sub-units

   1. ○ 5.1 Final Essay: When Deep Learning? Open this unit → ✓ ↻
   Open this module → Print certificate of completion