Machine Learning Experiments

Machine Learning Experiments is a browser based demo app that accompanies an open source GitHub repository of the same name by Oleksii Trekhleb. The repository holds the training code for each model, and the demo lets visitors try the trained models directly. It is aimed at students and developers who want to see how common neural network architectures behave on practical tasks. The demos include handwritten digit recognition and sketch recognition, each built with both a multilayer perceptron and a convolutional neural network. Visitors can play Rock Paper Scissors against the computer, using a convolutional network or MobileNetV2. Object detection runs on a camera stream with SSDLite and MobileNetV2 across 90 object categories from the COCO dataset, and image classification generates tags across 1,000 ImageNet categories. Recurrent neural network demos sum two numbers written as character sequences and generate Shakespeare style text, Wikipedia style text and cooking recipes with ingredients and instructions. A DCGAN demo generates clothing images from random noise. The demo is free to use and the site carries no pricing or paid plans. Anyone who wants to see how the models were trained is directed to the GitHub repository, which also accepts contributions.