New release of eScience Center’s explainable AI tool, DIANNA  

Version 1.7 of open-source Python package DIANNA (Deep Insight and Neural Network Analysis) has just been released. DIANNA is an explainable AI (XAI) tool that helps researchers get insights into the predictions of Deep Neural Networks. This new release of DIANNA now includes a vastly improved dashboard, extensive documentation, plus specific tutorials for earth science, astronomy and legal research. Watch this introductory video for a step-by-step demonstration of DIANNA. 

Elena Ranguelova, project leader of DIANNA and eScience Center’s Technology Lead for AI says, “The users of DIANNA can now not only obtain an intuitive understanding of how XAI works but also concrete guidelines on how to achieve explainability in three different scientific domains and types of data.”  

Presently, DIANNA is the only explainable AI tool supporting Open Neural Network Exchange (ONNX) – the de facto AI model standard that enables open science. Ranguelova says, “By supporting ONNX in DIANNA, we can wrap selected XAI methods, or explainers, in a simple, uniform interface. Although other open-source XAI libraries exist, they typically only support a single AI model format or data modality. For adoption by the wider community, an easy-to-use open-source tool facilitating the understanding of AI was needed.” 

In the new release of DIANNA, more tutorials and videos for three scientific domains have been added:  

  • Earth science: Model interpretation for terrestrial water-energy-carbon fluxes on tabular data 
  • Astronomy: Explainability of classifying radio-astronomy time-series as Fast Radio Bursts  
  • Legal research: Explaining a text classification of EU legislation sentences in regulatory or not 

Why was DIANNA developed? 

Most of today’s scientific challenges can only be tackled by using highly advanced technologies. Yet, many scientists are hesitant to adopt powerful AI models, due to a lack of trust and understanding of these models. XAI tools typically offer so-called explainers that provide insight into the AI’s predictions. They calculate the relevance of each input data item (image pixels, words in a text, etc.) to the final AI outcome. The resulting ‘relevance heatmaps’ indicate which pixels, words, signals, etc., contributed positively or negatively to the AI’s final decision.  

Overview of DIANNA features?  

  • Easy to use by non (X)AI experts; well-known XAI methods with a uniform interface 
  • Support for ONNX standard and converters from other formats 
  • Many data modalities: images, text, time series, and tabular 
  • A dashboard for comparison of different explainers for all data modalities  
  • Simple intuitive data, tasks and introductory tutorials to help you with your XAI understanding  
  • Several scientific use-cases tutorials to help you with your AI-enhanced research 

Visit DIANNA’s project page and software page on the Research Software Directory to learn more. 

Who are we?   

The eScience Center is a research organization dedicated to applying research software to answer research questions in any scientific domain through project collaborations. It has the largest concentration of dedicated, high-level research software expertise in the Netherlands. The eScience Center also has a Fellowship Programme and makes all of its software and training materials openly available online. For more information about what we offer, visit esciencecenter.nl.