Pooling layers (MaxPool, AvgPool) in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Pooling layers (MaxPool, AvgPool)

Problem:You have a convolutional neural network (CNN) trained on a small image dataset. The model currently uses only convolutional layers without any pooling layers.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Training loss: 0.15, Validation loss: 0.65

Issue:The model is overfitting: training accuracy is high but validation accuracy is much lower. The model may be memorizing training images and not generalizing well.

Your Task

Add pooling layers (MaxPooling and AveragePooling) to reduce overfitting and improve validation accuracy to at least 80%, while keeping training accuracy below 90%.

You must keep the convolutional layers but add pooling layers after them.

Do not change the dataset or increase its size.

Do not add dropout or other regularization methods in this experiment.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models

# Load example dataset
mnist = tf.keras.datasets.mnist
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# Normalize data
X_train, X_test = X_train / 255.0, X_test / 255.0

# Expand dims to add channel dimension
X_train = X_train[..., tf.newaxis]
X_test = X_test[..., tf.newaxis]

# Build CNN model with pooling layers
model = models.Sequential([
    layers.Conv2D(32, (3,3), activation='relu', input_shape=(28,28,1)),
    layers.MaxPooling2D((2,2)),
    layers.Conv2D(64, (3,3), activation='relu'),
    layers.AveragePooling2D((2,2)),
    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])

model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

history = model.fit(X_train, y_train, epochs=10, batch_size=64, validation_split=0.2)

# Evaluate on test data
test_loss, test_acc = model.evaluate(X_test, y_test)

print(f'Test accuracy: {test_acc:.4f}')

Added MaxPooling2D layer after the first convolutional layer to reduce spatial size and keep important features.

Added AveragePooling2D layer after the second convolutional layer to smooth features and reduce overfitting.

Kept convolutional layers and dense layers unchanged to maintain model capacity.

Results Interpretation

Before adding pooling layers: Training accuracy was 95%, validation accuracy was 70%, showing overfitting.

After adding pooling layers: Training accuracy dropped to 88%, validation accuracy improved to 82%, and test accuracy reached 81%. Loss values also decreased on validation data.

Pooling layers help reduce overfitting by shrinking feature maps and focusing on important features. This improves the model's ability to generalize to new data.

Bonus Experiment

Try replacing the AveragePooling layer with another MaxPooling layer and observe how validation accuracy changes.

💡 Hint

MaxPooling often preserves the strongest features better, which might improve or worsen validation accuracy depending on the dataset.

Practice

1. What is the main purpose of a MaxPooling layer in a neural network?

easy

A. To reduce the size of the input while keeping the strongest features

B. To increase the size of the input by adding more pixels

C. To randomly change pixel values for data augmentation

D. To convert images into grayscale

Solution

1. Step 1: Understand MaxPooling function

   MaxPooling scans small regions of the input and selects the maximum value, reducing size but keeping strong signals.

2. Step 2: Compare options with MaxPooling role

   Only To reduce the size of the input while keeping the strongest features describes reducing size while keeping strongest features, matching MaxPooling's purpose.

3. Final Answer:

   To reduce the size of the input while keeping the strongest features -> Option A

4. Quick Check:

   MaxPooling = reduce size + keep max [OK]

Hint: MaxPooling keeps max values, shrinking input size [OK]

Common Mistakes:

• Confusing MaxPooling with data augmentation
• Thinking MaxPooling increases input size
• Mixing MaxPooling with color conversion

2. Which of the following is the correct way to create a 2x2 MaxPooling layer with stride 2 in TensorFlow?

easy

A. tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=2)

B. tf.keras.layers.MaxPooling2D(pool_size=2, stride=2)

C. tf.keras.layers.MaxPooling2D(pool=(2, 2), strides=(2, 2))

D. tf.keras.layers.MaxPooling2D(size=2, stride=2)

Solution

1. Step 1: Recall correct TensorFlow MaxPooling2D syntax

   The correct parameters are pool_size and strides, both can be tuples or integers.

2. Step 2: Check each option's parameter names and values

   tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=2) uses pool_size=(2, 2) and strides=2, which is valid. Others use wrong parameter names or formats.

3. Final Answer:

   tf.keras.layers.MaxPooling2D(pool_size=(2, 2), strides=2) -> Option A

4. Quick Check:

   Correct params: pool_size, strides [OK]

Hint: Use pool_size and strides parameters exactly [OK]

Common Mistakes:

• Using 'stride' instead of 'strides'
• Using 'pool' or 'size' instead of 'pool_size'
• Passing integers without tuple when tuple needed

3. Given the input tensor below, what is the output shape after applying a 2x2 MaxPooling layer with stride 2?

input_tensor = tf.constant([[[[1], [2], [3], [4]],
                             [[5], [6], [7], [8]],
                             [[9], [10], [11], [12]],
                             [[13], [14], [15], [16]]]], dtype=tf.float32)

medium

A. (1, 3, 3, 1)

B. (1, 4, 4, 1)

C. (1, 2, 2, 1)

D. (1, 1, 1, 1)

Solution

1. Step 1: Understand input shape and pooling parameters

   Input shape is (1, 4, 4, 1). Pool size is 2x2, stride 2, so output spatial dims shrink by factor 2.

2. Step 2: Calculate output shape after pooling

   Height and width: floor(4/2) = 2. Channels and batch size remain same. Output shape is (1, 2, 2, 1).

3. Final Answer:

   (1, 2, 2, 1) -> Option C

4. Quick Check:

   Pooling halves spatial dims = (1, 2, 2, 1) [OK]

Hint: Output dims = input dims divided by stride [OK]

Common Mistakes:

• Forgetting batch and channel dimensions
• Using input size instead of divided size
• Confusing stride with pool size

4. You wrote this code but get an error:

layer = tf.keras.layers.AvgPooling2D(pool_size=2, strides=2)
output = layer(input_tensor)

What is the most likely cause of the error?

medium

A. input_tensor must be a numpy array, not a tensor

B. pool_size must be a tuple, not an integer

C. AvgPooling2D does not accept strides parameter

D. input_tensor is missing batch dimension

Solution

1. Step 1: Check input_tensor shape requirements

   Pooling layers expect 4D input: (batch, height, width, channels). Missing batch dimension causes error.

2. Step 2: Verify other options

   pool_size can be int or tuple, strides is accepted, and input can be tensor. So only missing batch dimension is likely cause.

3. Final Answer:

   input_tensor is missing batch dimension -> Option D

4. Quick Check:

   Pooling needs 4D input [OK]

Hint: Pooling layers require 4D input tensors [OK]

Common Mistakes:

• Passing 3D tensor without batch size
• Thinking strides parameter is invalid
• Confusing tensor and numpy array input

5. You want to reduce noise in feature maps but keep important edges. Which pooling strategy should you use and why? MaxPooling or AveragePooling?

hard

A. Use MaxPooling because it keeps strongest signals like edges

B. Use AveragePooling because it smooths out noise by averaging

C. Use MaxPooling because it averages values to reduce noise

D. Use AveragePooling because it selects max values to keep edges

Solution

1. Step 1: Understand effect of MaxPooling and AveragePooling

   MaxPooling keeps max values, emphasizing strong signals like edges but may keep noise spikes. AveragePooling smooths values, reducing noise.

2. Step 2: Match goal with pooling type

   To reduce noise but keep edges, smoothing by AveragePooling is better. MaxPooling keeps edges but may keep noise.

3. Final Answer:

   Use AveragePooling because it smooths out noise by averaging -> Option B

4. Quick Check:

   Noise reduction = AveragePooling [OK]

Hint: AveragePooling smooths noise; MaxPooling keeps edges [OK]

Common Mistakes:

• Confusing max with average effects
• Thinking MaxPooling reduces noise
• Mixing pooling effects on edges and noise