Machine Learning Systems (MLSysBook)
MLSysBook is an open, living textbook and hands-on learning stack that teaches AI engineering — the discipline of building efficient, reliable, safe, and robust intelligent systems that operate in the real world. The project is maintained by the Harvard EDGE / MLSysBook community and combines conceptual chapters with runnable code, educational frameworks, and hardware labs to bridge machine learning theory and systems practice.
Core components
- Textbook: an interactive online textbook covering foundations, design, performance, deployment, trust, and frontiers of ML systems. Chapters connect algorithms to systems-level tradeoffs (memory, latency, energy, accuracy, privacy, MLOps).
- TinyTorch: an educational, minimal deep learning framework included to teach internals by implementing autodiff, optimizers, CNNs, and transformers from scratch (code-licensed under Apache 2.0).
- Hardware kits & labs: exercises and kits for deploying models on edge devices (Arduino, Raspberry Pi, etc.) so learners face real constraints like memory, power, and timing.
- Co-labs & benchmarks: planned software co-labs and benchmarks for controlled experiments on latency, energy, and cost, enabling reproducible ML systems evaluation.
- Ecosystem & community: discussions, contribution guides, and an Open Collective fund to support outreach and hardware donations.
Why it’s useful
MLSysBook focuses on engineering AI systems—not just training isolated models. It teaches trade-offs between model accuracy and system constraints, gives reference implementations and reproducible benchmarks, and provides instructor- and learner-friendly materials for courses and self-study. The living nature of the book means it is updated by the community as practices and tools evolve.
Notable details
- License: textbook content under CC BY-NC-ND 4.0; TinyTorch code under Apache-2.0. This dual-license model separates educational content (share non-commercially, no derivatives) from modifiable software.
- Citation: a recommended IEEE citation (Reddi et al., CODES+ISSS 2024) is provided in the repo for academic referencing.
- Roadmap: hardcopy edition slated with MIT Press (noted as coming 2026), and additional co-labs and an "AI Olympics" are planned to expand hands-on and competitive learning.
How to get started
- Read the online textbook at the official site.
- Try introductory chapters (e.g., benchmarking, introduction) to learn core principles.
- Follow TinyTorch getting-started modules to implement learning building blocks.
- Run hardware labs to deploy models on edge devices and explore deployment trade-offs.
- Join Discussions or contribute via GitHub (issues, PRs, or contributions to TinyTorch and labs).
Intended audience
Students, instructors, researchers, and engineers who want to learn how to turn machine learning ideas into dependable systems suitable for real-world constraints. MLSysBook is particularly useful for courses on ML systems, MLOps, and embedded/edge ML.
Community & contribution
The repository lists many contributors and maintains contribution guidelines for the book, TinyTorch, and hardware labs. The project uses community feedback and reproducible reference implementations to close the loop between research and teaching.
