Subproject "Shared AI Theory"

The theory part of S3AI comprises mathematical aspects of transfer learning and a novel computational geometric approach for deep model analysis, e.g.

Tessellation: study the interdependence between a deep model represented as neural network, its induced (tessellation) geometry in the input space and its separability properties; Deep Transfer Learning: provide quantitative bounds on the misclassification and dertermine convergence rates in transfer learning settings;

Most results in machine learning are based on the assumption that the statistical data characteristics of the training setting and the application setting are the same. However, this assumption is often violated in industrial settings, especially, in the deep model sharing approach of S3AI. In our work, we provide algorithms, parameter choice methods and mathematical guarantees, to enable successful AI model adaptation as a core of the S3AI methodology. These findings have brought about follow-up research projects in industry 4.0 and personalized medicine:

1. In combination with Active Learning, to mitigate performance degradation in human-in-the-loop quality inspection due to changing influencing factories (FFG Project “FlexSpect.AI”, granted Oct 2021, lead: Profactor) with use cases in high-speed printing of packaging material and plastic moulding of 3D parts
2. In combination with federated learning, to lift the limited scope of applicability of microscopy-based histology analysis models from a local to a global scope; shipping a model trained in Europe to South Latin America can bring performance degradation due to the data shift caused by varying gene-expressivity in different populations; the S3AI approach enables the efficient retraining of such models (Vienna LifeScience Project “The Virtual Histopathologist”, granted February 2022, lead: TissueGnostics);