Skip to content

README.md

Latest commit

 

History

History
67 lines (47 loc) · 2.94 KB

README.md

File metadata and controls

67 lines (47 loc) · 2.94 KB

Deep Learning Mathematics Atlas (深度学习数学图鉴) 🚀

简体中文 | English

Build Status License: CC BY-NC-SA 4.0 Version LaTeX

Breaking the wall between Academic Math and Engineering Implementation.

This handbook is a "Rosetta Stone" for Deep Learning practitioners. It provides a direct, side-by-side mapping between rigorous mathematical definitions (LaTeX) and their actual implementations in modern frameworks (PyTorch).

🌟 Key Features

  • Rosetta Stone Layout: Side-by-side view of math formulas and PyTorch code.
  • SOTA Architectures: Detailed mathematical breakdown of Transformer, Mamba (SSM), LoRA, and Diffusion Models.
  • High-Quality Visuals: Standardized TikZ diagrams with professional aesthetics (shadows, semantic colors, gradient flow paths).
  • Mathematical Depth: Covers everything from Tensor basics and Linear Algebra to Stochastic Processes and Information Theory.
  • Academic Typography: Springer-level layout using Times New Roman and asymmetric wide margins for annotations.

📖 Content Structure

The handbook is organized into thematic "Parts":

  1. Foundations (基石篇): Tensors, Advanced Linear Algebra (SVD, QR), Probability Foundations.
  2. Anatomy (解剖篇): Activation Functions, Standard Layers (Linear, Conv), Sequence Layers (RNN, LSTM), Normalization.
  3. Objectives (目标篇): Distance Metrics, Information Theory & Probabilistic Losses.
  4. Dynamics (动力篇): Autograd (VJP/JVP), Optimization Algorithms, Stochastic Processes (Wiener, SDE).
  5. Architectures (架构篇): Multi-Head Attention, Generative Models (VAE, GAN).
  6. Foundation Models (大模型纪元): LLM Components, Mamba (SSM), PEFT (LoRA).
  7. The Frontiers (前沿探索): Adversarial Training, Graph Neural Networks, Quantization, Next-Gen Generative (Flow Matching).

🛠️ Build Instructions

Prerequisites

  • A TeX distribution: MiKTeX (Recommended for Windows) or TeX Live.
  • pdflatex compiler.

Build the PDF

We provide automation scripts for a seamless build:

Windows:

build.bat

macOS / Linux:

make

To clean intermediate files:

build.bat clean  # Windows
make clean       # Unix

🤝 Contributing

We welcome contributions to expand the "Atlas"! If you want to add a new SOTA model (e.g., DeepSeek MLA, FlashAttention) or fix a typo, please see CONTRIBUTING.md.

📜 License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Designed with ❤️ by Antigravity & Sisyphus.