core Estimated learning time: 7 h

6.12 Convolutional networks

You can build and train an image classifier from scratch.

Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing 11. The ECE Bridge — Signals, Edge & Embedded AI

Convolutions encode two priors — locality and translation equivariance — that make them the natural machine for images and one-dimensional signals, with receptive fields and parameter sharing explaining their efficiency. This topic builds and trains an image classifier from scratch. It opens the architectures run. The arithmetic deserves respect: stride, padding and dilation interact, and off-by-one shape errors are the standard tax paid by everyone who guesses instead of computes.

Work through these

Convolution, stride, padding, dilation
The operation that slides a small filter across an image, with settings controlling step size, edge handling and spacing. These four terms describe every convolution you will meet.
Pooling, receptive field, parameter sharing
Downsampling, the region of the input each output depends on, and reusing the same weights everywhere. Weight sharing is what makes these networks tractable on images.
Classic architectures: LeNet, AlexNet, VGG
Three historical architectures that introduced the ideas the rest of the field built on. Reading them in order is the fastest way to understand why modern designs look as they do.
1D convolutions for signals and time series
The same operation applied along one dimension works on signals and time series. This is the direct bridge from image work to the electronics and signals module.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.