Machine Learning Playground allows you to experiment with neural network models. The application ships with four read-only datasets. Some datasets are designed for regression analysis while others are suited for classification. The models themselves will consist of three layers: inputs, hidden and outputs. There is a set of hyper-parameters that you can adjust to gauge their effects on a models accuracy. You can define a stopping condition for halting the models training step via the number of epochs or the loss target value. You can also switch between two loss functions - Mean Squared Error and Cross-Entropy. Zoomable charts plot loss and accuracy across the epochs.
The app will also allow you to create and save custom datasets. You specify the number of inputs, hidden nodes and outputs along with the activation functions at the hidden and output layers. Tune your model built on your custom dataset using the hyper-parameters. The data files that you create are stored in the apps domain on the iPad. Those files will persist until you either delete them yourself using the delete feature on the Custom Datasets page, or you remove the app entirely from your iPad.