Report
‘Deep Learning Speedrun’ traces the math from real numbers to backpropagation
The article estimates its math prerequisites at half a freshman semester for someone “already really smart.”
TLDR
The author of ‘Deep Learning Speedrun’ asks what minimum math it takes to train an MNIST model. The article traces a path from real numbers and limits through derivatives to backpropagation and a small neural network. The author estimates the prerequisites at half a freshman semester for someone “already really smart,” but says the formal derivation takes about 3,000 lines of Lean and wonders whether it can be done faster.
Combined views
11.8K
3 Sources, first seen ago
