Machine Learning as a Reusable Microservice

Machine Learning is recently becoming a universal problem solving tool. However, implementing machine learning (ML) into applications is difficult, time intense, and requires expert knowledge. We encapsulate machine learning as a data-oriented microservice that can simply be used to mash up applications with machine learning capabilities. To illustrate the approach we identify three machine learning algorithms that are relevant for the Internet of Things (IoT): Feed-Forward Neural Networks (FFNN), Deep Believe Networks (DBN), and Recurrent Neural Networks (RNN). We analyze those algorithm's characteristic properties and model them as configurations for dynamically linkable REST ML service modules. Our approach strictly separates the algorithm implementation from its configuration. It allows a simple extension with diverse ML algorithms. Following a service oriented design, we implement the training of our neural networks as a separate module. We evaluate how the performance of our solution compares to directly programming the chosen TensorFlow library. Our approach facilitates the implementation of ML-based data analytics significantly by enabling reuse and sharing of executables and configurations. It enables rapid prototyping and an explorative use of ML.