Architecture Design (CNN under 25k params)

Conv1: 8 filters, 3×3 kernel, stride 1, padding 1 → output: 28×28×8

MaxPool: 2×2 → output: 14×14×8

Conv2: 16 filters, 3×3 kernel → output: 14×14×16

MaxPool: 2×2 → output: 7×7×16

Flatten: 7×7×16 = 784

FC: 784 → 10 classes

Parameter Count

Conv1: (3×3×1×8) + 8 = 80

Conv2: (3×3×8×16) + 16 = 1,168

FC: (784×10) + 10 = 7,850

Total = 9,098 parameters