Machine Learning Theory, 2,000 Slides Deep and Still Updating

What it does

This repository archives a continuously updated curriculum of machine learning notes—over 2,000 slides spanning foundational statistics, intermediate probabilistic models, and deep learning research topics. The material is delivered as PDF technical reports and accompanied by live bilingual seminars (Chinese Sunday streams and English Zoom sessions), plus links to older video tutorials and external MATLAB demos. It treats ML as an applied math discipline rather than a cookbook of frameworks.

The interesting bit

Most repos promise “notes” and dump a few Markdown files; this one is a longitudinal project actively replacing its own decade-old Beamer archive with new technical reports while running concurrent live classes on what the author calls “Gradient Descend Research.” The scope is deliberately anachronistic, pairing century-old regression theory with analyses of DeepSeek’s attention mechanisms in the same syllabus. That stubborn math-first approach is increasingly rare in an ecosystem optimized for quick API demos.

Key highlights

• Covers the full stack from model-evaluation bootstrapping and χ² tests to Rademacher complexity, PAC-Bayes, and Neural Tangent Kernels.
• Includes a dedicated Generative AI section analyzing Transformers with PyTorch code, KV caching, Decoupled RoPE, and DeepSeek’s Multi-Head Latent Attention.
• Research-grade deep learning topics: variance reduction (REBAR/RELAX), W-GAN and Bayesian GAN mathematics, Normalizing Flows, and NeuralODE.
• Links to live seminar schedules (Hong Kong time) and archived video recordings on Bilibili and YouTube.
• The author is actively validating and correcting ten years of prior notes, gradually replacing old Beamer slides with unified technical report PDFs.

Caveats

• The repository itself is primarily a collection of PDFs; executable code and Jupyter notebooks are largely absent here, with demos relegated to a separate MATLAB repository.
• The README is a dense index of links with minimal narrative structure, so navigating the 2,000+ slides requires knowing roughly what topic you are after.
• Some content is explicitly work-in-progress: the author notes they are still updating, validating, and correcting notes, and that certain lectures overlap across years (e.g., Softmax appears in both DeeCamp 2019 and Deep Learning Research Topics).

Verdict

Grab this if you are a researcher or advanced student who wants the mathematical scaffolding behind modern ML—particularly generative models and learning theory—and can tolerate PDFs over notebooks. Skip it if you are looking for copy-paste PyTorch scripts or a gentle, code-first introduction to the field.