// the find
trekhleb/machine-learning-experiments
🤖 Interactive Machine Learning experiments: 🏋️models training + 🎨models demo
A collection of ML experiments pairing Jupyter training notebooks with live browser demos, covering MLP, CNN, RNN, and GAN architectures using TensorFlow/Keras. Models are exported to TensorFlow.js and run entirely client-side. Aimed at developers learning ML fundamentals who want to see the full pipeline from training to inference.
The notebook-to-demo pairing is the actual value here — you can read the training code and then immediately interact with the result in the browser, which most learning resources don't bother doing. Shipping pre-converted TFjs models means zero backend required for the demos. The recipe generation RNN notebook is unusually detailed with a companion long-form writeup explaining the architecture decisions. Using MobileNetV2 transfer learning for rock-paper-scissors shows a realistic workflow, not just toy examples from scratch.
The whole repo is frozen in pre-Transformer ML — LSTM text generation and basic CNNs were already showing their age when this was published, and nothing here covers attention, transformers, or anything from the last few years of the field. Loading full model weights in the browser (tens to hundreds of MB per model) is called out in the README as bad practice, but the workaround is just 'this is an experiment' — there's no server inference path shown at all. The React demo app uses Flow types and a yarn 1.x lockfile, both of which are effectively abandoned tooling. Last meaningful commit activity appears to have stopped around 2021 despite a recent push date.