Deep Learning: Introduction to Convolutional Neural Networks (CNNs)

Introduction to Deep Learning

• Traditional machine learning algorithms rely on manual feature extraction from images by an expert, followed by training a classifier on these features.
• The performance of these algorithms depends heavily on the quality of the features previously found.
• In 2012, Convolutional Neural Networks (CNNs) broke records in the ILSVRC computer vision competition, revolutionizing image classification.

What is a Convolutional Neural Network (CNN)?

• CNNs are a subcategory of neural networks specifically designed to process input images.
• They consist of two main blocks: a feature extractor block and a classification block.
• The architecture of CNNs allows them to extract features of different complexities, from simple to sophisticated.

Convolution Layer

• The first block of a CNN functions as a feature extractor, applying convolutional filtering operations to the input image.
• The layer filters the image with multiple convolution kernels, returns feature maps, and normalizes/resizes them using an activation function.
• This process can be repeated multiple times, generating new feature maps that are filtered, normalized, and resized.

Pooling Layer and ReLU Correction Layer

• (No specific information provided in the text)

Fully Connected Layer

• The second block of a CNN transforms the input vector values using linear combinations and activation functions to return a new output vector.
• The output vector contains as many elements as there are classes, with each element representing the probability that the image belongs to that class.

Architecture of a CNN

• A CNN consists of four types of layers: convolution, pooling, ReLU correction, and fully-connected layers.
• The parameters of the layers are determined by backpropagation of the gradient, minimizing cross-entropy during the training phase.