Global average pooling

[cnuernber]

https://arxiv.org/pdf/1312.4400.pdf

@benkamphaus: Feel free to add any details we should consider when researching this feature.

[benkamphaus]

Follow on paper has a video demo-ing use for saliency: http://cnnlocalization.csail.mit.edu/

The big takeaways are:

• reduce huge expansion in parameters at final conv -> first fully connected layer
• reduce the capacity of that layer to overfit
• directly map from spatial features in a transparent way (saliency maps for free)

It's useful to compare how this works in keras here:

from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential

model = Sequential([
    Conv2D(48, kernel_size=(3, 3), input_shape=(224, 224, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    Conv2D(96, kernel_size=(3, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    Conv2D(128, kernel_size=(3, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    GlobalAveragePooling2D(),
    Dense(133, activation='softmax'),
])


model.summary()

Leads to:

Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_39 (Conv2D)           (None, 222, 222, 48)      1344      
_________________________________________________________________
max_pooling2d_37 (MaxPooling (None, 111, 111, 48)      0         
_________________________________________________________________
conv2d_40 (Conv2D)           (None, 109, 109, 96)      41568     
_________________________________________________________________
max_pooling2d_38 (MaxPooling (None, 54, 54, 96)        0         
_________________________________________________________________
conv2d_41 (Conv2D)           (None, 52, 52, 128)       110720    
_________________________________________________________________
max_pooling2d_39 (MaxPooling (None, 26, 26, 128)       0         
_________________________________________________________________
global_average_pooling2d_9 ( (None, 128)               0         
_________________________________________________________________
dense_10 (Dense)             (None, 133)               17157     
=================================================================
Total params: 170,789
Trainable params: 170,789
Non-trainable params: 0

Versus:

model = Sequential([
    Conv2D(48, kernel_size=(3, 3), input_shape=(224, 224, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    Conv2D(96, kernel_size=(3, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    Conv2D(128, kernel_size=(3, 3), activation='relu', kernel_initializer='he_normal'),
    MaxPooling2D((2, 2)),
    Flatten(),
    Dense(133, activation='softmax'),
])


model.summary()

With:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_42 (Conv2D)           (None, 222, 222, 48)      1344      
_________________________________________________________________
max_pooling2d_40 (MaxPooling (None, 111, 111, 48)      0         
_________________________________________________________________
conv2d_43 (Conv2D)           (None, 109, 109, 96)      41568     
_________________________________________________________________
max_pooling2d_41 (MaxPooling (None, 54, 54, 96)        0         
_________________________________________________________________
conv2d_44 (Conv2D)           (None, 52, 52, 128)       110720    
_________________________________________________________________
max_pooling2d_42 (MaxPooling (None, 26, 26, 128)       0         
_________________________________________________________________
flatten_4 (Flatten)          (None, 86528)             0         
_________________________________________________________________
dense_11 (Dense)             (None, 133)               11508357  
=================================================================
Total params: 11,661,989
Trainable params: 11,661,989
Non-trainable params: 0

That's: 11,661,989 vs: 170,789 trainable params! Dropping those and getting better performance (as detailed in the paper).