Pooling, Flattening & CNN Architecture

Pooling (Dimensionality Reduction)

After convolution, feature maps are still large. Pooling reduces size while keeping important info.

Max Pooling

Take the maximum value in each window:

Input (4×4):
[1 2 | 3 4]
[5 6 | 7 8]
-----+-----
[9 10| 11 12]
[13 14| 15 16]

Max Pool (2×2 window):
[6  8]     ← Max of [1,2,5,6], [3,4,7,8], etc.
[14 16]

Why? Keeps strongest activation (most relevant feature).

Average Pooling

Take average instead of max:

[1 2 | 3 4]   Average Pool (2×2):
[5 6 | 7 8]   [3.5  6.5]
-----+-----
[9 10| 11 12] [11.5 14.5]
[13 14| 15 16]

Full CNN Architecture

Input Image (224×224×3)
      ↓
Conv (64 filters, 3×3, ReLU) → 224×224×64
      ↓
MaxPool (2×2) → 112×112×64  (reduced!)
      ↓
Conv (128 filters) → 112×112×128
      ↓
MaxPool (2×2) → 56×56×128
      ↓
Conv (256 filters) → 56×56×256
      ↓
MaxPool (2×2) → 28×28×256
      ↓
Flatten → 200,704 values
      ↓
Dense (512, ReLU) → 512
      ↓
Dropout (0.5) → Drop half randomly
      ↓
Dense (1000, Softmax) → Final output

Common CNN Architectures

Use pre-trained: Don't train CNNs from scratch! Load pre-trained weights (ImageNet).

Why CNNs Work

1. Local connectivity: Only nearby pixels connected
2. Weight sharing: Same filter across all positions
3. Hierarchical learning: Build up from edges to objects
4. Translation invariance: Same object detected regardless of position