Why does Global Average Pooling replace the FC head in modern CNNs?
mediumAnswer
- Take the mean of each feature map across H, W to get a C-dim vector, then apply a linear classifier.
- Advantages: (1) drastic parameter reduction — no huge FC on flattened features; (2) intrinsic regularization; (3) enables class activation maps (CAM) since one channel per class is aligned with a class score; (4) works with any input resolution.
- Introduced in Network-in-Network and popularized by ResNet.
Check yourself — multiple choice
- Adds many parameters
- Average each feature map → C-dim vector → linear head → far fewer params, regularizes, enables CAM
- Only for RNNs
- GAP replaces the softmax
GAP: per-channel spatial average → few params, regularizes, class activation maps.
#cnn#architectures
Practise Deep Learning
214 interview questions in this topic.