Dendra
Watch it learn.
What it is
Dendra is a multilayer perceptron built from scratch, training live in the browser on XOR and concentric circles so you can watch the decision boundary form. It implements backpropagation with tanh/ReLU activations, Xavier/He initialization and full-batch gradient descent, and its 11 tests centre on a finite-difference gradient check. Dependency-light, free to download and run.
A multilayer perceptron with backpropagation (tanh/ReLU, Xavier/He init, full-batch gradient descent), training live on XOR and concentric circles. 11 tests, centred on a finite-difference gradient check.
What's inside
The full source, the tests, and CI. Open it, read it, change it. A zero-dependency core, free, in the MIT spirit.
Run it after unzip
pnpm install && pnpm dev