Video by freeCodeCamp.org · taught by Aakash N S · streamed from YouTube. All rights remain with the creator.
Lesson 5: Image Classification with Convolutional Networks
About this lesson
This fourth session of the Deep Learning with PyTorch live course, taught by Aakash N S, introduces convolutional neural networks, the architecture behind modern computer vision. It moves from greyscale digits to colour photographs, working with CIFAR-10, a dataset of small images of everyday objects and animals in ten classes. The session shows how to download the dataset, how three-channel RGB images are represented as tensors, how to explore the images and classes, and how to create training and validation sets and data loaders for them. It then explains what a convolution is: how a small grid of weights called a kernel slides across an image, multiplying and adding pixel values to produce a feature map, and how padding, stride and pooling affect the size of the output. You see why convolutions suit images so well, using few parameters and detecting the same pattern anywhere in a picture. A convolutional network is then built by stacking convolutional layers, activation functions and pooling layers, followed by fully connected layers that produce one score per class, and trained on a GPU. The session ends by discussing underfitting and overfitting, showing how to recognise each from the training and validation curves, and outlining ways to improve the model's performance, which the next session puts into practice. Each step is shown running live in a notebook. Afterwards, read Chapter 5 of the course book, which counts every parameter of a convolutional network by hand.