An interactive neural network powered digit classifier.
Seeing the accuracy on neural networks I was training is useful, but I figured it would be nice to be able to draw digits directly into a neural network and see live predictions.
So I trained a simple ANN with Tensorflow locally, then uploaded the weights to a web app with a canvas that lets you draw digits and view live predictions throughout the process.
Play around with it below! It's interesting to get a feel for what types of features the model is looking for with certain digits.
UI and drawing pad built with raw JavaScript using my own UI library canvasLib.
Neural network was trained locally with Python and Tensorflow on the MNIST dataset of handwritten digits. Model is then run in the browser using Tensorflow.js.