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.
- 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, 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.
The unified app is in unified-app/.
cd unified-app
npm install
npm run devBuild and test:
cd unified-app
npm test
npm run audit:quality
npm run test:smoke
npm run buildThe app uses React, Vite, Tailwind CSS, Three.js, GSAP, and Recharts.
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.
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, and fine-tuning.
The frontier path covers modern architecture and systems topics: dense vs MoE models, MLA and 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.
The retrieval track covers the RAG pipeline as a system: chunking, embedding search, vector indexing, reranking, context packing, grounding, retrieval metrics, and failure modes.
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.
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.
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.
The RL track covers agents, rewards, discounted returns, MDPs, value iteration, policy iteration, Q-learning, exploration, policy gradients, actor-critic methods, and reward shaping.
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/
Each directory has its own README with setup notes and examples.
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.
node scripts/deploy-github-pages.mjsThe 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.
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
MIT. See LICENSE.