Pooling Layers in R

Pooling layers are part of convolutional neural networks (CNNs). This article explores the concept of pooling layers, their types, and how to implement them in R.

Prerequisites: Convolutional Neural Network, Pooling Layers

Pooling layers perform down-sampling operations on feature maps, which:

1. Reduce Spatial Dimensions: This decreases computational requirements and memory usage.
2. Extract Dominant Features: By focusing on the most relevant features, pooling layers make models more robust to variations like translation or scaling.
3. Prevent Overfitting: By reducing the complexity of feature maps, pooling layers help the model generalize better.

The types of pooling layer includes Max pooling, Average Pooling and Global Pooling.

When to Use Pooling?

Pooling layers are typically used after convolutional layers to progressively reduce the spatial size of the feature maps. This hierarchical reduction is critical for deep architectures, enabling them to extract higher-level abstractions efficiently.

Before diving into the implementation, you may want to explore convolutional neural networks in R to understand the process better.

Creating a Model with Pooling Layers in R

To use pooling layers in R, we rely on the keras package, which provides a high-level interface to TensorFlow.

Setting Up the Environment

Ensure you have keras installed. If not, install it using:

install.packages("keras")
keras::install_keras()

Max Pooling in R

Max pooling layer extracts the maximum value from a region (e.g., 2x2 or 3x3) of the feature map. In R, we use maximum_pool_2d() in Keras to perform max pooling.

# Load the keras library
library(keras)

# Create a simple model with max pooling
model <- keras_model_sequential() %>%
  layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu", input_shape = c(64, 64, 3)) %>%
  layer_max_pooling_2d(pool_size = c(2, 2)) %>%  # Max Pooling layer
  layer_flatten() %>%
  layer_dense(units = 64, activation = "relu") %>%
  layer_dense(units = 10, activation = "softmax")

# Summary of the model
summary(model)

Output:

Model: "sequential"
_____________________________________________
 Layer (type)                       Output Shape                    Param #     
==================================================
 max_pooling2d (MaxPooling2D)       (None, 31, 31, 32)              0           
 flatten (Flatten)                  (None, 30752)                   0           
 dense_1 (Dense)                    (None, 64)                      1968192     
 dense (Dense)                      (None, 10)                      650         
==================================================
Total params: 1,969,738
Trainable params: 1,969,738
Non-trainable params: 0
_____________________________________________

Average Pooling in R

Average pooling layer computes the average of the values in a region. In R, layer_average_pooling_2d() computes averages within a patch.

# Create a simple model with average pooling
model <- keras_model_sequential() %>%
  layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu", input_shape = c(64, 64, 3)) %>%
  layer_average_pooling_2d(pool_size = c(2, 2)) %>%  # Average Pooling layer
  layer_flatten() %>%
  layer_dense(units = 64, activation = "relu") %>%
  layer_dense(units = 10, activation = "softmax")

# Summary of the model
summary(model)

Output:

Model: "sequential_1"
____________________________________________
 Layer (type)                       Output Shape                    Param #     
=================================================
 conv2d_1 (Conv2D)                  (None, 62, 62, 32)              896         
 average_pooling2d (AveragePooling2  (None, 31, 31, 32)             0           
 D)                                                                             
 flatten_1 (Flatten)                (None, 30752)                   0           
 dense_3 (Dense)                    (None, 64)                      1968192     
 dense_2 (Dense)                    (None, 10)                      650         
=================================================
Total params: 1,969,738
Trainable params: 1,969,738
Non-trainable params: 0
____________________________________________

Global Pooling R

Global Max Pooling extracts the maximum value from the entire feature map using layer_global_max_pooling_2d(). Global Average Pooling averages all feature map values into one value using layer_average_pooling_2d.

# Global Max Pooling example
model <- keras_model_sequential() %>%
  layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu", input_shape = c(64, 64, 3)) %>%
  layer_global_max_pooling_2d() %>%  # Global Max Pooling
  layer_dense(units = 64, activation = "relu") %>%
  layer_dense(units = 10, activation = "softmax")
summary(model)

# Global Average Pooling example
model <- keras_model_sequential() %>%
  layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu", input_shape = c(64, 64, 3)) %>%
  layer_global_average_pooling_2d() %>%  # Global Average Pooling
  layer_dense(units = 64, activation = "relu") %>%
  layer_dense(units = 10, activation = "softmax")

# Summary of the model
summary(model)

Output:

Model: "sequential_6"
_____________________________________________
 Layer (type)                       Output Shape                    Param #     
==================================================
 conv2d_8 (Conv2D)                  (None, 62, 62, 32)              896         
 global_max_pooling2d_1 (GlobalMaxP  (None, 32)                     0           
 ooling2D)                                                                      
 dense_13 (Dense)                   (None, 64)                      2112        
 dense_12 (Dense)                   (None, 10)                      650         
==================================================
Total params: 3,658
Trainable params: 3,658
Non-trainable params: 0
_____________________________________________

Model: "sequential_7"
______________________________________________
 Layer (type)                       Output Shape                    Param #     
===================================================
 conv2d_9 (Conv2D)                  (None, 62, 62, 32)              896         
 global_average_pooling2d_1 (Global  (None, 32)                     0           
 AveragePooling2D)                                                              
 dense_15 (Dense)                   (None, 64)                      2112        
 dense_14 (Dense)                   (None, 10)                      650         
===================================================
Total params: 3,658
Trainable params: 3,658
Non-trainable params: 0
_______________________________________________

Key Parameters for Pooling Layers

1. Pool Size: The dimensions of the pooling window (e.g., 2x2 or 3x3).
2. Strides: The step size for moving the pooling window. If not specified, it defaults to the pool size.
3. Padding: Whether to use "valid" (no padding) or "same" (pads input to ensure output size matches input size).

Benefits of Using Pooling Layers

• Reduced Model Complexity: Pooling layers simplify feature maps, speeding up training and inference.
• Improved Generalization: By removing less relevant details, pooling makes the model more robust.
• Invariant Representations: Helps models ignore small spatial transformations in input data.

Pooling layers helps to balance computational efficiency with feature extraction. With R and the keras package, implementing these layers is straightforward, enabling R users to build powerful deep learning models for image analysis and beyond.