Why CNNs understand visual patterns in TensorFlow - Challenge Your Understanding

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why do convolutional layers use small filters?

Convolutional Neural Networks (CNNs) use filters (kernels) to scan images. Why do these filters usually have small sizes like 3x3 or 5x5?

A Small filters capture local patterns and reduce the number of parameters, making learning efficient.

B Small filters increase the image size to help the network see more details.

C Small filters are used to convert images into grayscale before processing.

D Small filters randomly remove pixels to reduce noise in images.

💡 Hint

Think about how images have small details like edges and textures.

Small filters focus on local features like edges or textures. This helps the CNN learn important visual patterns efficiently without too many parameters.

❓ Predict Output

intermediate

Output shape after convolution

Given a 28x28 grayscale image input to a convolutional layer with 16 filters of size 3x3, stride 1, and 'valid' padding, what is the output shape?

TensorFlow

import tensorflow as tf
input_shape = (1, 28, 28, 1)  # batch size 1, height 28, width 28, channels 1
inputs = tf.random.normal(input_shape)
conv_layer = tf.keras.layers.Conv2D(filters=16, kernel_size=3, strides=1, padding='valid')
outputs = conv_layer(inputs)
print(outputs.shape)

A (1, 27, 27, 16)

B (1, 28, 28, 16)

C (1, 25, 25, 16)

D (1, 26, 26, 16)

💡 Hint

Use formula: output_size = input_size - filter_size + 1 for 'valid' padding.

With 'valid' padding, output height and width = 28 - 3 + 1 = 26. Number of filters = 16 channels.

❓ Model Choice

advanced

Choosing CNN architecture for texture recognition

You want to build a CNN to recognize textures in images. Which architecture choice helps the network better capture fine texture details?

A Use dropout layers only without convolution to prevent overfitting.

B Use a single large 11x11 filter in the first layer to capture all texture at once.

C Use deeper layers with small 3x3 filters and max pooling to gradually capture complex patterns.

D Use only fully connected layers without convolution to focus on global features.

💡 Hint

Think about how stacking small filters helps build complex features step-by-step.

Stacking small filters in deeper layers helps the CNN learn simple to complex texture features gradually, improving recognition.

❓ Metrics

advanced

Interpreting CNN training loss and accuracy

During CNN training on image classification, you observe the training loss steadily decreases but validation accuracy plateaus early. What does this indicate?

A The model has perfect generalization and no further improvement is possible.

B The model is overfitting the training data and not generalizing well to new images.

C The model is underfitting and needs more training epochs.

D The training data is corrupted causing loss to decrease but accuracy to plateau.

💡 Hint

Think about what it means when training improves but validation does not.

Decreasing training loss with stagnant validation accuracy usually means the model memorizes training data but fails to generalize, a sign of overfitting.

🔧 Debug

expert

Why does this CNN fail to learn meaningful features?

Consider this TensorFlow CNN code snippet. After training, the model accuracy stays near random guessing. What is the main issue?

TensorFlow

import tensorflow as tf
model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(32, kernel_size=3, activation='relu', input_shape=(28,28,1)),
    tf.keras.layers.MaxPooling2D(pool_size=2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(10, activation='softmax')
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
# Training code omitted

A The model lacks normalization layers like BatchNormalization, causing unstable training.

B The model uses 'relu' activation which is not suitable for CNNs.

C The input shape is incorrect; it should be (28,28) without channel dimension.

D The loss function 'sparse_categorical_crossentropy' is incompatible with softmax activation.

💡 Hint

Think about what helps CNNs train stably on image data.

Without normalization layers, CNN training can be unstable or slow, leading to poor learning and low accuracy.