Deep Learning, No Compiler Required

What it does This repository distributes Excel workbooks that reconstruct machine-learning algorithms—Softmax, backpropagation, LSTMs, ResNets, Transformers, Mamba, even AlphaFold—using only spreadsheet formulas. Each file breaks the architecture into cells you can inspect, edit, and trace by hand.

The interesting bit By trapping state-of-the-art models inside a tool built for expense reports, the project forces every operation into explicit, auditable arithmetic. The resulting friction is the feature: you cannot hide behind import torch when the entire forward pass is visible in cell H14.

Key highlights

• Span from workbook fundamentals (Dot Product, Matrix Multiplication, Linear Layer) to advanced architectures (xLSTM, Self-Attention, Multihead Attention).
• Ships a blank DeepSeek worksheet covering Multi-head Latent Attention and Mixture of Experts.
• Runs anywhere Excel (or a compatible spreadsheet app) does; no Python environment or GPU required.
• Roadmap includes GAN, VAE, U-Net, and CLIP additions.

Caveats

• The DeepSeek lecture is currently blank only, so it is a skeleton rather than a completed exercise.
• The README is essentially a directory listing; it does not describe how the sheets are structured or what prior knowledge they assume.

Verdict Ideal for practitioners who understand AI conceptually but want to touch the actual math. Less useful if you are hunting for training scripts or production-ready implementations.