Why CNNs understand visual patterns in TensorFlow - Experiment to Prove It

Experiment - Why CNNs understand visual patterns

Problem:We want to understand why Convolutional Neural Networks (CNNs) are good at recognizing visual patterns like edges and shapes in images.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%

Issue:The model is overfitting, learning training images too well but not generalizing to new images.

Your Task

Reduce overfitting so validation accuracy improves to at least 85% while keeping training accuracy below 92%.

Keep the CNN architecture simple (2 convolutional layers).

Do not increase training data size.

Use TensorFlow and Keras only.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Load dataset
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.cifar10.load_data()

# Normalize images
train_images, test_images = train_images / 255.0, test_images / 255.0

# Data augmentation
augmenter = ImageDataGenerator(
    rotation_range=15,
    width_shift_range=0.1,
    height_shift_range=0.1,
    horizontal_flip=True
)

# Build CNN model
model = models.Sequential([
    layers.Conv2D(32, (3,3), activation='relu', input_shape=(32,32,3)),
    layers.BatchNormalization(),
    layers.MaxPooling2D((2,2)),
    layers.Dropout(0.25),

    layers.Conv2D(64, (3,3), activation='relu'),
    layers.BatchNormalization(),
    layers.MaxPooling2D((2,2)),
    layers.Dropout(0.25),

    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dropout(0.5),
    layers.Dense(10, activation='softmax')
])

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

# Train model with augmentation
history = model.fit(
    augmenter.flow(train_images, train_labels, batch_size=64),
    epochs=30,
    validation_data=(test_images, test_labels)
)

# Output final metrics
train_acc = history.history['accuracy'][-1] * 100
val_acc = history.history['val_accuracy'][-1] * 100
print(f'Training accuracy: {train_acc:.2f}%')
print(f'Validation accuracy: {val_acc:.2f}%')

Added dropout layers after convolution and dense layers to reduce overfitting.

Added batch normalization after convolution layers to stabilize training.

Used data augmentation to create varied training images.

Kept CNN architecture simple with two convolutional layers.

Results Interpretation

Before: Training accuracy was 95%, validation accuracy was 70%. The model memorized training images but failed on new images.

After: Training accuracy dropped to 90%, validation accuracy improved to 86%. The model learned general visual patterns better.

Adding dropout, batch normalization, and data augmentation helps CNNs avoid overfitting and better understand visual patterns that generalize to new images.

Bonus Experiment

Try increasing the number of convolutional layers to 4 and observe if validation accuracy improves further.

💡 Hint

More layers can learn more complex patterns but may also increase overfitting. Use dropout and batch normalization carefully.

Practice

1. Why do Convolutional Neural Networks (CNNs) understand visual patterns well?

easy

A. Because they use random guessing to classify images

B. Because they detect simple shapes and combine them to form complex patterns

C. Because they memorize every pixel in the image

D. Because they ignore the image details and focus on text data

Solution

1. Step 1: Understand CNN feature detection

   CNNs use filters to detect simple shapes like edges and corners in early layers.

2. Step 2: Combine features for complex patterns

   Later layers combine these simple shapes to recognize bigger and more complex visual patterns.

3. Final Answer:

   Because they detect simple shapes and combine them to form complex patterns -> Option B

4. Quick Check:

   Simple shapes + combination = CNN understanding [OK]

Hint: Remember CNNs build complex patterns from simple shapes [OK]

Common Mistakes:

• Thinking CNNs memorize entire images
• Believing CNNs guess randomly
• Assuming CNNs ignore image details

2. Which TensorFlow layer is used to detect features like edges in images?

easy

A. Conv2D

B. Dense

C. Dropout

D. Flatten

Solution

1. Step 1: Identify feature detection layer

   The Conv2D layer applies filters to images to detect features like edges and textures.

2. Step 2: Compare with other layers

   Dense is for fully connected layers, Dropout prevents overfitting, Flatten reshapes data, so they don't detect features.

3. Final Answer:

   Conv2D -> Option A

4. Quick Check:

   Conv2D detects features [OK]

Hint: Conv2D = convolutional filters for images [OK]

Common Mistakes:

• Confusing Dense with Conv2D
• Using Dropout for feature detection
• Thinking Flatten detects features

3. What will be the output shape after applying a Conv2D layer with 32 filters and kernel size (3,3) on input shape (28, 28, 1) with default padding in TensorFlow?

medium

A. (28, 28, 1)

B. (28, 28, 32)

C. (30, 30, 32)

D. (26, 26, 32)

Solution

1. Step 1: Understand Conv2D output shape with default 'valid' padding

   With kernel size 3 and input 28x28, output size = 28 - 3 + 1 = 26 for width and height.

2. Step 2: Account for number of filters

   The output depth equals the number of filters, which is 32.

3. Final Answer:

   (26, 26, 32) -> Option D

4. Quick Check:

   Output shape = (input - kernel + 1, filters) [OK]

Hint: Output size = input - kernel + 1 with 'valid' padding [OK]

Common Mistakes:

• Assuming output shape equals input shape
• Confusing padding types
• Ignoring number of filters in output depth

4. Identify the error in this TensorFlow CNN code snippet:

model = tf.keras.Sequential([
  tf.keras.layers.Conv2D(16, (3,3), activation='relu', input_shape=(28,28,1)),
  tf.keras.layers.MaxPooling2D(pool_size=2),
  tf.keras.layers.Dense(10, activation='softmax')
])

medium

A. Missing Flatten layer before Dense layer

B. Wrong activation function in Conv2D

C. Input shape should be (28,28) not (28,28,1)

D. MaxPooling2D should come after Dense layer

Solution

1. Step 1: Check layer order and compatibility

   Conv2D and MaxPooling2D output 3D tensors, but Dense expects 1D input.

2. Step 2: Identify missing Flatten layer

   Flatten layer is needed to convert 3D output to 1D before Dense layer.

3. Final Answer:

   Missing Flatten layer before Dense layer -> Option A

4. Quick Check:

   Flatten needed before Dense [OK]

Hint: Add Flatten before Dense after Conv/Pooling layers [OK]

Common Mistakes:

• Skipping Flatten layer
• Changing activation unnecessarily
• Misplacing MaxPooling2D

5. You want a CNN to recognize handwritten digits. Which combination best helps the model focus on important visual details and reduce image size?

hard

A. Dense layers followed by Dropout layers

B. Flatten layer followed by Conv2D layers

C. Conv2D layers followed by MaxPooling2D layers

D. MaxPooling2D layers followed by Dense layers

Solution

1. Step 1: Understand Conv2D role

   Conv2D layers detect visual features like edges and shapes in images.

2. Step 2: Understand MaxPooling2D role

   MaxPooling2D reduces image size by keeping important features, helping focus and reduce computation.

3. Final Answer:

   Conv2D layers followed by MaxPooling2D layers -> Option C

4. Quick Check:

   Conv2D + MaxPooling = feature focus + size reduction [OK]

Hint: Use Conv2D then MaxPooling for image features and size [OK]

Common Mistakes:

• Using Dense before Conv2D
• Flatten before Conv2D
• Ignoring pooling layers