The website displays a network consisting of nodes and edges. As the neural network trains on the MNIST handwritten digit dataset in real-time, the displayed network is dynamically updated such that it models changes in the neural network. The site also displays a line plot that tracks error over the training process and heatmaps that display each pixel’s impact on the selected neuron.
The Multilayer Perceptron(MLP) used in the site was implemented from scratch in Javascript(based on Michael Nielsen's implementation in Neural Networks and Deep Learning) and uses stochastic gradient descent as well as minibatching.