Teachable Machine.

The model learns from your examples

Teachable Machine offers a visual introduction to machine learning. You collect examples, assign classes, start training and test new inputs. Official materials describe image, sound and pose recognition. The task is to distinguish the classes you prepared, rather than provide general conversation or generate illustrations.

The project belongs to Google Creative Lab’s experiments. Its open repository publishes libraries, examples and an explanation of the TensorFlow.js architecture. It also warns that the experiment is not an official Google production product with guaranteed support. That limitation matters more for a learning project than the developer’s familiar name.

Why change the background when checking

Suppose you want to distinguish pencils from spoons in camera images. Collect your own examples with varied object positions, distances and lighting. Do not photograph one class only on white backgrounds and the other only on dark backgrounds: the model may learn the background rather than the objects.

After training, use new images and change the capture conditions separately. Test an absent target object and an object from no known class. A high score on a familiar image says little about another room. Held-out images should represent the conditions where recognition will actually be used.

Export and local work

The official repository explains that the training and execution of models occur in the browser. Google Drive can be used separately to save the project. Exported models can be connected to sites and applications; examples for multiple languages and supporting libraries are given. This is not a promise of a ready-made cloud API for any task.

When integrating a model into your own project, check uncertain predictions and new input conditions. Consequential decisions need stricter data and error controls than a learning demonstration. Open code explains the mechanism, but component licenses and the final application’s reliability still need review.

Avoid collecting only near-identical neighboring frames. Show the object from different angles and test it outside its familiar setting. If the application will encounter unknown objects, define what counts as an uncertain response and what users should see. Otherwise, the classifier must choose a known class even for unsuitable input.

Suitable tasks

• Train an educational visual classifier
• Export a model into your own small project