Module 06 Activity

Scenario

Build an image classifier and evaluate it the way an imbalanced real dataset demands.

Show that a convolution gives the same response to a pattern in two different corners.
Compute output sizes for four convolution configurations using the size rule, then verify each.
Trace shapes layer by layer through a CNN and measure the flatten width rather than deriving it.
Normalise a batch per channel and demonstrate a flip is reversible.
Evaluate on an imbalanced three-class problem. Report accuracy, then per-class recall.
Rank errors by confidence and list the three most confident mistakes.

A classifier report: architecture, parameter count, accuracy, majority baseline, confusion matrix, and your three worst errors.

out = floor((in + 2*padding - kernel) / stride) + 1. Kernel 3 with padding 1 preserves size.
Two poolings take 28 to 7. With 16 channels the flatten width is 784.
On the imbalanced example, accuracy looks acceptable while one class has recall 0.00.

Read per-class recall before accuracy. Accuracy is a summary of the majority class.