A comparison of open-source real-time online learning frameworks for neural networks

With the increasing amount of data and the requirement to derive insights from this data, it is important to have the ability to update the models that are used with new data in real-time. This paper provides an analysis of existing solutions for real-time online machine learning using neural networks. Six different machine learning frameworks are examined. Quantitative analysis is performed on the following metrics: samples per second, predictions per second, the time between new data and inclusion in the model. Performance evaluation is carried out using a benchmarking framework that is created for this purpose and released as open-source software. In the qualitative comparison the following aspects are evaluated: preprocessing, normalization, hardware support, drift countermeasures, ensemble learning support and license.