DS 6050: Deep Learning — Course Website

The public website for DS 6050 Deep Learning (School of Data Science, University of Virginia), taught by Heman Shakeri. Twelve modules take you from linear models and backpropagation to transformers, parameter-efficient fine-tuning, and generative AI — through video lectures, readings from the open Dive into Deep Learning textbook, Colab notebooks, and self-check questions.

Live site: https://shakeri-lab.github.io/dl-course-site/ YouTube playlist: https://www.youtube.com/playlist?list=PLwZv_wzaKu3vzW88m6ASKemCTb70ukycH

The site is fully public and self-paced. For-credit logistics (deadlines, quizzes, submissions, grades) live in UVA's Canvas, not here.

Stack

• Next.js 15 (App Router, static export) + React 19 + TypeScript
• Tailwind CSS + a trimmed set of shadcn/ui components
• KaTeX for math in self-check questions (rendered at build time)
• Deployed to GitHub Pages by .github/workflows/deploy.yml on every push to main

Where content lives

All course content is data, not markup. To update content you almost never touch JSX:

File	Contents
lib/module-data.ts	Per-module lectures (YouTube IDs, slide PDFs), Colab links, readings, descriptions
lib/module-extras.ts	Per-module learning objectives, time estimates, prerequisites, D2L section links, homepage topics
lib/self-check-data.ts	Per-module self-check questions (KaTeX $...$ math supported)
lib/site-config.ts	Site-wide facts: URLs, instructor, playlist, analytics code, "content current as of"
public/	Slide PDFs (Modules 11–12), syllabus PDF, logo, social-share card, llms.txt

Pages (app/…/page.tsx) render that data through components/module-template.tsx and friends. SEO infrastructure (sitemap, robots, JSON-LD, Open Graph) is generated from the same data — adding a module to module-data.ts updates everything automatically.

Common edits

• New lecture video: add { title, videoId } to the module's lectures array in lib/module-data.ts.
• Change a reading: edit d2lLinks in lib/module-extras.ts (each entry is a verified https://d2l.ai/... section URL).
• Add a self-check question: append to the module's array in lib/self-check-data.ts.
• Enable analytics: create a free GoatCounter site and put its code in siteConfig.goatcounter.
• New semester: update siteConfig.contentCurrentAsOf, tag the release (git tag spring-2027), adjust content as needed.

Development

npm install
npm run dev        # http://localhost:3000
npm run build      # static export to out/ (what GitHub Pages serves)

The production build serves under the /dl-course-site base path (see next.config.mjs); local dev runs at the root.

Deployment

Push to main → GitHub Actions builds the static export and publishes it to GitHub Pages. Nothing manual. A separate weekly workflow (link-check.yml) sweeps all external links (D2L, arXiv, YouTube) and opens an issue if any rot.

Design principles

• One learning path per module — watch → read → self-check → code, all on the module page. D2L is the canonical text for Modules 1–9; for Modules 10–12 (ViT, PEFT, quantization, multimodal, diffusion) the lecture slides and linked papers are the readings, because the book doesn't cover them at course depth.
• No slide duplication — lectures are board-first; slides are published only where they substitute for not-yet-recorded videos (Modules 11–12). Keeping slides out of the site elsewhere avoids drift against the LaTeX sources.
• Evergreen — no dates on the public site; semester logistics stay in Canvas.

License

Original course materials are licensed CC BY 4.0. Lecture videos © Heman Shakeri. Linked readings (papers, D2L) remain under their own licenses.