Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Phase 2 — Deep Learning

This is where neural networks arrive, and where Modelwright’s experience-first principle does its most important work. You will train a real neural network with PyTorch and watch it learn before you build any of the machinery underneath it. Only then, knowing exactly what a network is and what it is for, do you open it up and construct the engine that powers it, from scratch.

That ordering matters. Building an automatic differentiation engine is one of the most clarifying things you can do in all of machine learning, but it is a miserable place to start. By the time you reach it here, you will have used a network, seen gradients do their work, and have real questions that the from-scratch build answers. The build stops being abstract and becomes the satisfying reveal of how something you already use actually works.

From there the phase climbs the architecture ladder: tensors and a full network from scratch, then real command of PyTorch, the craft of making training actually work, and the major architectures (convolutional networks, sequence models, and finally the transformer), each used first and then built.

Modules (work them in order)

Phase capstone

Capstone 2: reproduce a published result. Take a small model or a compact training run, reproduce it from scratch, and write up where your numbers matched, where they did not, and why. Ship it to its own repository, modelwright-phase-2-capstone.

Begin with Module 2.1.