← Back to writing

Series

math-behind-neural-networks

6 parts, in order.

  1. 01

    Aug 18, 2026

    The Math Behind Neural Networks, Part 1.4 — Which Loss to Choose, and Why

    The output distribution was a choice. Other choices give other costs, each predicting a different statistic — and why cross-entropy beats squared error when the gradient would otherwise fall silent.

  2. 02

    Aug 2, 2026

    The Math Behind Neural Networks, Part 1.3 — Where the Loss Function Comes From

    Cost functions derived from maximum likelihood. Why probabilities multiply, how the logarithm keeps the order— the top-down route to the loss we assumed in 1.1.

  3. 03

    Jul 25, 2026

    The Math Behind Neural Networks, Part 0 — Where Deep Learning Comes From

    What a neural network is before any of the math: why deep learning had to exist, what every term actually means, and why a model must learn its own features. The one model the whole field grows from, built from zero.

  4. 04

    Jun 9, 2026

    The Leash That Makes AI Polite

    How a 1951 statistics formula quietly keeps your chatbot from going feral. KL divergence, RLHF, and the tension between them.

  5. 05

    May 24, 2026

    The Math Behind Neural Networks, Part 1.2 — The Mathematics of Learning

    The derivative, derived by hand from MSE, and gradient descent built on top of it. Where the 2/n comes from, why we subtract the gradient, and a full worked training example.

  6. 06

    May 9, 2026

    The Math Behind Neural Networks, Part 1.1 — The Mathematical Foundations

    Slope, linear regression, and the error function, built from the ground up. Why we divide by n, why we square the error, and why a loss value alone cannot tell you how to improve.