The IJCAI International Joint Conference on Artificial Intelligence is one of the top 10 conferences worldwide in the field of AI and takes place annually alternating on a different continent, this year from July 23rd to 29th, 2022 in Vienna. This was an ideal opportunity for SCCH to present itself in the context of our S3AI COMET module to the more than 2,000 participants with a booth in the Exhibition Hall and to position itself with expertise in the area of AI.
The focus was on Graph Neural Networks, Causal Reasoning and also neuro-symbolic AI with applications for AI-assisted programming.
S3AI General Chair Bernhard Moser presented a result of the project's ongoing research on Logic-Geometry Decomposition of Neural Networks entitled "Tessellation-Filtering ReLU Neural Networks" in front of about 300 interested people. The presented methodology allows to reconstruct the geometry of the decision space in the high-dimensional input space from the hard-to-interpret neural network parameters and is thus fundamental for the comparability and analysis of neural networks.

As part of the S3AI project, our colleague Michal Lewandowski attended the ELLIS Machine Learning Summer School, which took place at the University of Cambridge from 11th to 15th July. The Summer School covered a wide range of topics in the field of AI, e.g. Graph Neural Networks, Bayesian NNs, or Casual Inference, to mention a few. Michal is a Researcher and Data Scientist at SCCH, and showed in his poster presentation a measure that reveals a local topological structure of ReLU Neural Networks, but not only. „We propose a way to evaluate models beyond accuracy. We conjecture, that our measure will help to indicate a neural network, which will generalize better to unseen data”, says Michal. "Around 130 participants from all over the world took part in the Summer School and so there was good networking and I had a wonderful time, and I'm overwhelmingly grateful for the experience and the opportunity to meet so many great researchers“, concludes Michal.