Phase 4 · Deep learning · Architectures
Convolutional networks
Pattern detectors for images.
Open this lesson Create a free account
Why it matters
Images have structure a plain MLP ignores: nearby pixels belong together, and a cat is a cat wherever it appears in the frame. Convolutional networks build that in, which is why they powered the deep-learning breakthrough in computer vision in 2012. They're still inside many image systems, and the same ideas (sliding windows, shared weights) appear in audio models and in parts of multimodal LLMs.
What you'll be able to do
- Explain convolution: a small filter of shared weights sliding over an image
- Compute output sizes from kernel size, stride and padding
- Build a convolution from patches and a matrix multiply, and train a small CNN
- Compare a CNN's parameter efficiency with a fully connected network, and know where CNNs are used today
Before this lesson
Next in this phase: Recurrent networks