Machine Learning on Smartphones and Smartwatches: Next Word Prediction with TensorFlow
Recently, machine learning has been increasingly used not only on the server side, but also in client applications designed for end users. Often these applications have to run on mobile phones, tablets or even smart watches – low-performance devices that do not always have a good Internet connection. Thus, the models should not only predict accurate results and work offline, but also take into account the limitations of the devices on which they will operate.
In this talk we will share our experience of using Recurrent Neural Networks for language modeling, implementing our model in TensorFlow for building smart keyboard predictions and integrating them into Key2 Swipe application on Android and iOS smartphones and smart watches – a keyboard with special layout optimized for small screens.