Interactive machine learning using BIDMach

Machine learning is growing in importance in industry, the sciences, and many other fields. In many and perhaps most of these applications, users need to trade off competing goals and build different model prototypes rapidly, which requires much human intelligence and is time consuming. Therefore, interactive customization and optimization aims to help expert incorporate secondary criteria into the model-generation process in an interactive way. In this paper we describe the design of BIDMach machine learning system, with an emphasis of performing customized and interactive model optimization. The keys of the design are (i) a machine learning architecture which is modular, and support primary and secondary loss functions (ii) high-performance training so that non-trivial models can be trained in real-time (using roofline design and GPU hardware) (iii) highly-interactive visualization tools that support dynamic creation of visualizations and controls to match the bespoke criteria being optimized. Also, when we later turn to deep neural networks which require hours or days for training, we discuss how our current framework could be extended to support monitoring and optimizing learning schedule. 1 BIDMach: high-performance, customized machine learning framework Figure 1: BIDMach’s Architecture Visualization in Browser Web Server Computing Engine Grab data from GPU 3~5 times/s User input Model parameters Using D3.js Through WebSocket Manipulate parameters Figure 2: Visualization Architecture The first key to interactive, customized machine learning is an architecture which supports it. BIDMach [5] is a new machine learning toolkit which has demonstrated extremely high performance with modest hardware (single computers with GPUs), and which has the modular design shown in Fig 1. BIDMach uses minibatch updates, typically many per second, so that models are being updated continuously. This is a good match to interactive modeling, since the effects of analysts actions will be seen quickly. Rather than a single model class, models comprise first a primary model, which typically outputs the model loss on a minibatch and a derivative or other update (Gibbs sampling for example) for it. Next an optimizer is responsible for updating the model given gradients. Finally, “mixin” represents secondary constraints or likelihoods. Primary models and mixins are combined