Build a Brain in 10 Minutes

Recognizing an image with an MLP is possible but inefficient. CNNs exploit the spatial structure of pixels: the same filter slides across the whole image, weights are shared, and accuracy jumps from 97% to 99% on MNIST — and from unusable to industrial-grade on ImageNet.

The 2017 pivot: drop recurrence, do everything with attention. We cover embeddings, the attention mechanism, and close the course with a capstone that pulls together everything you've learned.