Grant: Alliance Grants

One major hurdle to the adoption of Artificial Intelligence (AI) lies in its reliance labeled data. Real-world datasets may be missing labels required for supervised learning. They may exhibit data imbalance, making it difficult for AI models to generalize to rare events, or have biases, which lead to downstream predictions that lack fairness. For instance, identifying rare but crucial events, such as malicious attacks or novel events, remains a significant challenge due to the scarcity of labeled instances. Proposed solutions to these challenges, such as data augmentation or data synthesis can introduce potential biases, which lead to failure to generalize or even worse unfair predictions. Similarly, self-supervised or un-supervised learning and reduce the reliance on labeled data, but require vast amount of data, which complicates fairness and safety guarantees. Furthermore, real-world data can exhibit distribution drift, which needs to be identified and excluded as either out-of-distribution, or trigger model updates that require even more labeled training data. Finally, the use of AI in decision making is increasingly under regulator scrutiny, and rightly so, as these technologies, while useful, are relatively new and untested. To address these challenges, we propose a Research Community focussed on An End-to-End Approach to Safe and Secure AI Systems. The explicit goal of this research community will be to research, develop and demonstrate mechanisms and techniques to solve the multitude of challenges with the use of AI outlined above. Specifically, our goals are: 1. Develop methods that will enable the training of AI models in scenarios that lack reliable, labeled data without having to directly rely on external, untrusted pre-trained foundation models. 2. Develop techniques to ensure the security, robustness, fairness and interpretability of AI models. 3. Develop guard rails for AI use to both ensure regulator compliance