Handwritten digit recognition runs in Scratch

The creator frames the project as a counter to the assumption that machine learning work happens in Python. Scratch is known as a visual language for beginners and for ambitious builds like 3D games, and this experiment places a fully connected neural network in that same environment.

The model uses 784x255=199920, a shape the creator says looks odd because it was fitted to Scratch's list limit of roughly 200,000 items. That constraint was later relaxed: after learning sprites could be generated automatically, the creator never tested whether 784x256=200704 would also work.

The digit recognizer is shared as 文字認識.sb3, and Python users can train the model and auto-generate the sprite locally or on Google Colab. Accuracy is described as neither good nor bad, and the creator notes the network's speed may be a separate question from whether chatbots could be built this way.