https://github.com/danielsobrado/machine-learning-visualized
Learn Machine learning by doing exercises and intuitive animations
https://github.com/danielsobrado/machine-learning-visualized
ai-agents attention-mechanism gradient-descent hands-on-machine-learning learning-by-doing llms lstm-neural-networks matrix-multiplication neural-networks stable-diffusion transformers vae-implementation
Last synced: 11 days ago
JSON representation
Learn Machine learning by doing exercises and intuitive animations
- Host: GitHub
- URL: https://github.com/danielsobrado/machine-learning-visualized
- Owner: danielsobrado
- Created: 2025-11-28T04:53:57.000Z (7 months ago)
- Default Branch: main
- Last Pushed: 2026-05-26T05:38:39.000Z (16 days ago)
- Last Synced: 2026-05-26T07:19:21.777Z (16 days ago)
- Topics: ai-agents, attention-mechanism, gradient-descent, hands-on-machine-learning, learning-by-doing, llms, lstm-neural-networks, matrix-multiplication, neural-networks, stable-diffusion, transformers, vae-implementation
- Language: JavaScript
- Homepage: https://danielsobrado.github.io/Machine-Learning-Visualized/
- Size: 56 MB
- Stars: 6
- Watchers: 0
- Forks: 2
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- Changelog: change-of-basis-animation/index.html
- Agents: AGENTS.md
Awesome Lists containing this project
README
# Machine Learning Visualized
Machine Learning Visualized is an interactive curriculum for machine learning, deep learning, language models, retrieval, diffusion, reinforcement learning, and the math behind them.
The project started as a collection of standalone animations. It is now centered on a unified React app with guided paths, lesson metadata, quizzes, labs, glossary links, and local progress tracking.
[Open the live site](https://danielsobrado.github.io/Machine-Learning-Visualized/)
## What is inside
- A unified lesson browser with searchable topics and curriculum tracks.
- Guided paths for fundamentals, experimentation and causal ML, LLMs, frontier LLMs and agentic systems, RAG, model reliability, vision and diffusion, and reinforcement learning.
- Core ML lessons for splitting data, cross-validation, leakage, scaling, metrics, calibration, PCA, clustering, tree ensembles, and classical classifiers.
- Model reliability lessons for debugging, interpretability, monitoring, fairness, and uncertainty estimation.
- Experimentation and causal ML lessons for A/B testing foundations and power analysis, with planned modules for sequential testing, CUPED, confounding, DAGs, treatment effects, and propensity scores.
- Transformer lessons for attention, masks, architecture families, training objectives, token generation, sampling, KV cache, Flash Attention, and fine-tuning.
- Frontier LLM lessons for MoE at scale, MLA, Native Sparse Attention, RLVR/GRPO, test-time compute, long-context systems, omni multimodal models, diffusion language models, efficient serving, frontier evaluation/safety, tool-using reasoners, and agentic coding systems.
- RAG lessons for chunking, vector indexing, reranking, grounding, retrieval evaluation, and failure modes.
- Neural-network lessons for backpropagation, initialization, optimizers, dropout, batch normalization, and training-loop dynamics.
- Diffusion lessons from beginner denoising intuition through sampling, classifier-free guidance, U-Net vs DiT, SD3, DiT, VAE, CLIP, T5, and flow matching.
- Small from-scratch implementations in Rust, Go, Java, and Python for neural networks, diffusion, and Markov chains.
## Current App
The unified app is in `unified-app/`.
```bash
cd unified-app
npm install
npm run dev
```
Build and test:
```bash
cd unified-app
npm test
npm run audit:quality
npm run test:smoke
npm run build
```
The app uses React, Vite, Tailwind CSS, Three.js, GSAP, and Recharts.
## Screenshots
### Core ML Lesson
### LLM Generation Lesson
### Frontier LLM Architecture
### Reasoning RLVR / GRPO
### Efficient LLM Serving
### Frontier Evaluation and Safety
### Diffusion Basics Lesson
## Curriculum Areas
### Foundations
The foundations track covers linear algebra, probability, statistics, optimization, and the core supervised-learning workflow. Lessons include matrix multiplication, linear regression, train/validation/test splits, gradient descent, PCA, k-means, overfitting, regularization, calibration, ROC and precision-recall curves, and bias-variance tradeoffs.
### Natural Language Processing and Transformers
The NLP and transformer track starts with bag-of-words, tokenization, and embeddings, then moves into attention, self-attention, masks, positional encoding, RoPE, transformer architectures, LLM training objectives, token generation, sampling, KV cache, Flash Attention, Native Sparse Attention, and fine-tuning.
### Frontier LLMs and Agentic Systems
The frontier path covers modern architecture and systems topics: dense vs MoE models, MLA, Native Sparse Attention, attention compression, reasoning models, RLVR/GRPO, test-time compute, tool-using reasoning, agentic coding, long-context systems, omni multimodal models, diffusion language models, efficient LLM serving, and frontier evaluation/safety.
### RAG
The retrieval track covers the RAG pipeline as a system: chunking, embedding search, vector indexing, reranking, context packing, grounding, retrieval metrics, and failure modes.
### Model Reliability
The model reliability track covers post-training and deployed-system concerns: debugging failures, interpreting model behavior, estimating uncertainty, monitoring drift and regressions, and evaluating fairness tradeoffs across slices and groups.
### Experimentation and Causal ML
The experimentation track connects hypothesis testing, confidence intervals, metrics, calibration, leakage, fairness, monitoring, and uncertainty to causal decision-making. Active lessons now cover A/B testing foundations, power and sample size, sequential testing and peeking, CUPED variance reduction, confounding and Simpson's paradox, causal graphs and DAGs, treatment effects, and propensity scores.
The next-priority applied ML pillars are also active as overview lessons: time series and forecasting, recommender systems and ranking, ML security and robustness, efficient inference and compression, and data engineering for ML.
### Vision and Diffusion
The diffusion track starts with basic denoising and sampling before moving into classifier-free guidance, U-Net vs DiT, latent VAEs, CLIP, T5, SD3, DiT, joint attention, and flow matching.
### Reinforcement Learning
The RL track covers agents, rewards, discounted returns, MDPs, value iteration, policy iteration, Q-learning, exploration, policy gradients, actor-critic methods, and reward shaping.
## Standalone Implementations
The repository also includes compact implementations meant for reading and experimentation:
- `mini-nn/`, `mini-nn-go/`, `mini-nn-java/`, `mini-nn-python/`
- `mini-diffusion/`, `mini-diffusion-go/`, `mini-diffusion-java/`, `mini-diffusion-python/`
- `mini-markov/`, `mini-markov-go/`, `mini-markov-java/`, `mini-markov-python/`
- `mini-eagle/` Rustlings-style exercises for EAGLE 3.1 speculative decoding
- `mini-spec-sparse/` Rustlings-style exercises for SpecSA / SpecAttn sparse speculative decoding
- `mini-turboquant/` Rustlings-style exercises for TurboQuant KV-cache quantization
Each directory has its own README with setup notes and examples.
## Publishing
GitHub Pages is published manually from this machine. The deploy script builds the unified app and pushes the generated site to the `gh-pages` branch.
```bash
node scripts/deploy-github-pages.mjs
```
The script also publishes static `*-animation/index.html` entry pages with route-specific metadata, so older animation URLs and crawlers land on the current unified lessons.
## Repository Layout
```text
unified-app/ Unified React app
screenshots/readme/ Current README screenshots
scripts/ Local maintenance and deploy scripts
*-animation/ Static lesson entry pages and legacy standalone lessons
mini-nn*/ Small neural-network implementations
mini-diffusion*/ Small diffusion implementations
mini-markov*/ Small Markov-chain implementations
mini-eagle/ Speculative decoding exercises with TODO-driven Rust tests
mini-spec-sparse/ Sparse speculative decoding exercises with TODO-driven Rust tests
mini-turboquant/ KV-cache quantization exercises with TODO-driven Rust tests
```
## License
MIT. See [LICENSE](LICENSE).