Teachable Machine

This Teachable Machine review covers what the web-based tool does, its features, cost, how to train a model step by step, and where it fits next to other machine learning tools. Teachable Machine is a Google Creative Lab experiment: you show the computer examples, it learns to recognize patterns, and you test the model live with your webcam or microphone. Instead of being explicitly programmed, the model learns from the examples you give it. Version 2.0, launched in November 2019, added sound and pose projects to the original image classifier from 2017.

What is Teachable Machine used for?

Teachable Machine is used to create machine learning models for classification: telling apart images, sounds or body poses. Typical projects sort objects in front of a webcam, detect simple sounds such as a clap or a whistle, recognize hand gestures or body movements, or tell a cat from a dog. Teachers use it in lessons to show students how artificial intelligence learns from data, and makers use it to control games, websites and physical projects such as an Arduino robot.

Because it is a web based tool with no install, anyone with a browser has access to it, which makes it easy to build projects with friends or a class. It is also a research tool. Google's 2020 paper reports that over 182,000 users in 201 countries created more than 125,000 classification models with it, and that places such as Stanford d.school, NYU ITP and the MIT Media Lab used it in teaching (Google Research).

Key features

• Three project types for custom machine learning models: image, audio and pose projects, each trained on examples you record with a webcam or microphone or upload as files.
• Training in the browser: training runs on your own computer, so the data stays on the device unless you save the project to Google Drive (Google blog).
• Live testing: the preview shows the model's guess and its confidence for each class as you move or make sounds.
• Export: models export as TensorFlow.js for websites, and image models also as TensorFlow and TensorFlow Lite, including an Edge TPU build for Coral boards (GitHub issue #66).
• Shareable link: you can upload a model and load it from a hosted URL with the open-source libraries, which include code snippets for JavaScript, Java and Python (GitHub).

Teachable Machine pricing

Teachable Machine is free to use. There is no account, subscription or paid plan: you open the site and start creating. Saving projects to Google Drive needs a Google account. The helper libraries are open source under the Apache 2.0 license. The cost to check is support: Google describes the project as "an experiment, not an official Google product" and promises only to do its best to maintain it (GitHub).

Step by step guide: train a machine learning model

1. Open the web-based tool and choose an image, audio or pose project.
2. Name each class, such as "thumbs up" and "thumbs down".
3. Add examples to every class with the webcam, the microphone or uploaded pictures. More varied examples give better results.
4. Click Train model. Training runs in the browser.
5. Test the model in the preview and add examples where it guesses wrong.
6. Export the model or upload it to get a shareable link for your apps and sites.

What happens after you upload images: they are used as training examples in your browser session and are not sent to Google unless you save the project to Drive or upload the trained model.

Teachable Machine for kids and classrooms

Teachable Machine is popular with kids and teachers because it turns abstract concepts into play. Students collect their own input, such as pictures, poses or sounds from musical instruments, train a model and see at once how the technology guesses. That makes it easy to explain how teaching computers differs from programming them, and to evaluate why a model gets something wrong. Lessons often start with a fun idea, such as a rock-paper-scissors game or a gesture-controlled game, and end with a discussion of bias in the training data. Google's design team describes the goal as training by example rather than rule-making (Google Design).

Where Teachable Machine shines — and where it falls short

Strengths:

• Free, fast and easy: a first model takes minutes, with no coding.
• Data stays in the browser by default.
• Good for lessons, prototypes and creative experiments.
• Exports ML models to web, mobile and edge devices, so you can easily train a model and use it in a complete project.

Trade-offs:

• Classification only; no object detection, regression or text models.
• Small datasets and simple models; not built for production accuracy.
• An experiment with best-effort support, not an official Google product.
• Limited control over training settings compared with a real ML platform.

How Teachable Machine compares

• Create ML: Apple's no-code model trainer for Mac, with more model types and Core ML output for Apple apps.
• H2O.ai: open-source and enterprise AutoML for tabular data and production models.
• DataRobot: enterprise AutoML and model operations for business data.
• Obviously AI: no-code predictions on spreadsheet data; the company has since rebranded as Zams.

Is Teachable Machine worth it?

Yes, for learning, teaching and quick prototypes. It is the fastest free way to show students or colleagues how a computer learns to recognize images, sounds and poses, and the export options are enough for a web demo or a Coral project. It is not a platform for production machine learning: models are small, support is best effort, and more complex tasks need a real ML tool. If you want to turn a Teachable Machine prototype into a product, low-code agencies and developers on LowCodeDevs can help build it.

Sources

• Teachable Machine 2.0, Google blog: launch, project types, in-browser training and privacy.
• Teachable Machine community repository, GitHub: libraries, snippets, license and experiment status.
• Teachable Machine paper, Google Research: design and usage figures.
• Designing a Teachable Machine, Google Design: training by example and design choices.
• Teachable Machine, Experiments with Google: the project page by Google Creative Lab.
• Export formats, GitHub issue #66: TensorFlow, TF Lite and Edge TPU exports for image models.