Currently, there is an emerging paradigm shift in artificial intelligence to support a collaborative deep learning environment in which new and pre-trained deep models are pooled, reused and composed in a reconfigurable manner. This development has versatile effects on the scope of AI applications: quantitatively, in terms of reducing development costs by reusing pre-trained models and saving data acquisition efforts; qualitatively, by overcoming limitations in the availability of annotated data, e.g., for small lot size problems.
However, there are major challenges to be addressed: a) security, in terms of protection against adversarial attacks and in terms of preserving privacy; b) safety, by understanding the model’s behavior and providing guarantee key figures.
These challenges are to be tackled in three aspects:
- different information baseline scenarios such as domain adaptation, multi-task learning, multi-view learning (information fusion);
- different modes of collaboration such as sharing data (no model), sharing representations (partial model) and sharing models (full model) or model components;
- different classes of deep learning strategies such as CNN, LSTM, Capsule Networks, GANs;
The project is primarily oriented by fundamental research that aims to come up with
- a mathematical framework for regularizing the hidden representations of the deep models during the learning phase to allow the derivation of error bounds and functionality guarantees;
- robust transfer learning techniques to accelerate training;
- distributed network architectures and learning schemes that guarantee privacy preserving;
- analysis methodologies for assuring security and safety.
S3AI is a COMET Module within the COMET – Competence Centers for Excellent Technologies Programme and funded by BMK, BMDW and the government of Upper Austria.
The COMET Programme is managed by FFG.