Machine Learning in the Browser

Abstract

The past decade has seen the rise of rich, dynamic Web applications, and it has also seen the popularization of machine learning. Despite this, we have not seen web applications that evaluate machine learning models in the browser because of technical limitations that make it difficult to do so quickly. I motivate and present a machine learning library for the web, which interoperates with one of the most popular machine learning libraries and is capable of evaluating such models quickly. I further discuss the new classes of features and products enabled by such a library, including: privacy, offline mode, and self-contained demos. The library runs within an order of magnitude of the speed that the native, single-threaded equivalent runs in. For example, one popular image recognition model, Inception, ran in 0.67 seconds natively and as fast as 2.59s on commodity browsers by leveraging emerging web technologies. I also discuss the challenges of building and some of the use cases of such a library.