We are delighted to announce that multiple innovations developed within the EVENFLOW project have been recognised by the European Commission’s Innovation Radar. This development highlights the Toolkit for Scalable Online Training and Incremental Model Construction, a breakthrough innovation designed by the “Athena” Research Center (ARC) to advance adaptive, online neural learning over streaming data. In addition, the Verification Toolkit for learning-based CEF by Imperial College of Science, Technology and Medicine and the Reasoning-based Forecast Interpreter by NCSR Demokritos have also been acknowledged, further showcasing the project’s impact and contributions to cutting-edge research and innovation.

About EVENFLOW innovations

Toolkit for Scalable Online Training and Incremental Model Construction: To achieve its goals, the Scalability toolkit  consists of (i) a sophisticated data synopses engine implemented on Apache Flink using a novel Synopses-Data-Engine-as-a-Service paradigm, capable of summarising voluminous, rapid streams that are fed to the online training process of neural and neuro symbolic models, (ii)  new synopses-based training techniques which utilise the generated synopses to achieve appropriate balance between training accuracy and training time, preserving the real-timeliness of the involved streaming applications, (iii) a novel suite of data-driven synchronisation protocols for data-parallel learning, orthogonal to (i) and (ii), to further scale out the training process combining high accuracy with reduced training time in parallel training settings. In that, in EVENFLOW AI-powered learning and forecasting techniques meet the requirements of scalable processing over Big Data, combining the virtues of both worlds.

Verification Toolkit for learning-based CEF: The verification toolkit for online neuro-symbolic learning for Complex Event Forecasting (CEF). A verification engine for CEF, implementing novel EVENFLOW methods for the verification of robustness of neural classifiers. It consists of novel methods for verifying the robustness of neural event predictors, addressing important challenges such as the high dimensionality and the temporal dependencies of the input streams, the complexity of the resulting neural predictors and the non-linearity of the activation functions.

Reasoning-based forecast interpreter: Reasoning-based interpreter of forecasts, produced by the learnt models. It will be based on online neuro-symbolic learning and reasoning techniques, tailored to dynamic, stream-based applications. The neural and the symbolic parts will be continuously co-evolving, assisting each other towards robust and accurate forecasting.