Skip to content

Convolutional Neural Network (CNN) - Notes

Table of Contents (ToC)

Introduction

A Convolutional Neural Network (CNN) is a deep learning model particularly effective for image and spatial data processing.

What's a Convolutional Neural Network (CNN)?

  • A type of deep neural network designed to process structured grid data such as images.
  • Utilizes convolutional layers to automatically detect spatial hierarchies in data.
  • Primarily used in image recognition, object detection, and related fields.

Key Concepts and Terminology

  • Convolutional Layer: Applies filters to input data to create feature maps.
  • Pooling Layer: Reduces the spatial dimensions of feature maps, retaining essential information.
  • ReLU (Rectified Linear Unit): Activation function used to introduce non-linearity.
  • Fully Connected Layer: Connects neurons from one layer to another, typically used before the output layer.

Applications

  • Image Classification: Identifying objects or subjects in images.
  • Object Detection: Locating and identifying objects within images.
  • Facial Recognition: Matching faces in images to identities.
  • Medical Imaging: Analyzing and diagnosing conditions from medical images (e.g., X-rays, MRIs).

Fundamentals

CNN Architecture Pipeline

  • Input Layer: Accepts image data typically in 2D (height x width x channels).
  • Convolutional Layer: Applies filters to detect features like edges, textures, and patterns.
  • Activation Function (ReLU): Applies non-linearity after each convolution operation.
  • Pooling Layer: Downsamples feature maps to reduce dimensionality and computation.
  • Fully Connected Layer: Combines features for final classification or regression tasks.
  • Output Layer: Provides the final predictions, often using softmax for classification.

How CNNs Work

  • Convolution Operation: Uses filters (kernels) to slide over input data and compute dot products, generating feature maps.
  • Feature Extraction: Early layers capture low-level features (edges), while deeper layers capture high-level features (shapes, objects).
  • Pooling: Reduces spatial dimensions to keep computation efficient and to prevent overfitting.
  • Flattening: Converts 2D feature maps into a 1D vector for input into fully connected layers.
  • Classification/Regression: The final fully connected layers output predictions based on extracted features.

Types of CNN Architectures

  • LeNet-5: Early CNN architecture used for digit recognition.
  • AlexNet: Introduced deeper networks with more filters and layers, won the 2012 ImageNet competition.
  • VGGNet: Uses very small (3x3) convolution filters but with a deep network architecture.
  • ResNet (Residual Networks): Introduced skip connections to allow for very deep networks, solving the vanishing gradient problem.
  • Inception Network (GoogLeNet): Combines convolutions of different sizes in parallel to capture varying spatial features.

Some Hands-On Examples

  • Image Classification: Using CNNs to classify CIFAR-10 or MNIST datasets.
  • Object Detection: Implementing YOLO or Faster R-CNN for detecting objects in images.
  • Facial Recognition: Building a CNN model to recognize faces from a dataset.
  • Medical Image Analysis: Applying CNNs to classify diseases in medical scans.

Tools & Frameworks

  • TensorFlow: Popular framework for building and training CNN models.
  • Keras: High-level API for easy implementation of CNNs, built on top of TensorFlow.
  • PyTorch: Widely used for flexible CNN development with dynamic computation graphs.
  • OpenCV: Library for image processing, often used alongside CNNs for pre-processing.

Hello World!

from tensorflow.keras import layers, models

# Build a simple CNN model
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.MaxPooling2D((2, 2)),
    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])

# Example model summary
model.summary()

Lab: Zero to Hero Projects

  • Digit Recognizer: Build a CNN to classify handwritten digits using the MNIST dataset.
  • Dog vs Cat Classifier: Train a CNN to distinguish between images of dogs and cats.
  • Real-Time Object Detection: Implement a CNN-based object detection system using YOLO.
  • Medical Image Classification: Develop a CNN to diagnose diseases from X-ray images.

References

  • LeCun, Y., et al. (1998). "Gradient-based learning applied to document recognition."
  • Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). "ImageNet Classification with Deep Convolutional Neural Networks."
  • Simonyan, K., & Zisserman, A. (2014). "Very Deep Convolutional Networks for Large-Scale Image Recognition."
  • He, K., et al. (2015). "Deep Residual Learning for Image Recognition."

  • Szegedy, C., et al. (2015). "Going Deeper with Convolutions."

  • CNNs: https://en.wikipedia.org/wiki/Convolutional_neural_network

  • CNNs HF Course: https://huggingface.co/learn/computer-vision-course/unit2/cnns/introduction
  • CNN Models: CNN Notes
  • Lab: CNN Notebooks