Towards Trustworthy AI via Federated and Explainable Machine Learning

Our group works to advance the frontiers of trustworthy artificial intelligence (AI). We focus on the fundamental limits and practical methods of federated and explainable machine learning. Several grants support these efforts:

  • A Mathematical Theory of Trustworthy Federated Learning (MATHFUL)
    Funded by the Research Council of Finland (Decision #363624)
    More Info

  • A Mathematical Theory of Federated Learning (TRUST-FELT)
    Funded by the Jane and Aatos Erkko Foundation (Decision #A835)
    More Info

  • Forward-Looking AI Governance in Banking & Insurance (FLAIG)
    Supported by Business Finland

Federated Learning: Computational and Statistical Innovations

Federated Learning for Network-Structured Data

The era of big data over networks demands solutions for handling interconnected, heterogeneous datasets. A prime example is pandemic management, where wearables and smartphones generate local datasets tied by physical, social, or biological network structures. These datasets often vary in statistical properties but exhibit intrinsic relationships.

Our work introduces networked exponential families, a probabilistic model that enables:

  • Adaptive Data Pooling: Groups similar datasets to build personalized predictions.
  • Privacy-Preserving Techniques: Ensures data confidentiality in federated settings.
  • Scalable Optimization: Employs robust methods to handle imperfections in data and computation.

  • “Clustered Federated Learning via Generalized Total Variation Minimization”
    Y. SarcheshmehPour, Y. Tian, L. Zhang, A. Jung, IEEE Transactions on Signal Processing, 2023.
    doi:10.1109/TSP.2023.3322848

  • “On the Duality Between Network Flows and Network Lasso”
    A. Jung, IEEE Signal Processing Letters, 2020.
    doi:10.1109/LSP.2020.2998400

  • “Networked Exponential Families for Big Data Over Networks”
    A. Jung, IEEE Access, 2020.
    doi:10.1109/ACCESS.2020.3033817

  • “Semi-Supervised Learning in Network-Structured Data via Total Variation Minimization”
    A. Jung et al., IEEE Transactions on Signal Processing, 2019.
    doi:10.1109/TSP.2019.2953593

  • “Localized Linear Regression in Networked Data”
    A. Jung and N. Tran, IEEE Signal Processing Letters, 2019.
    doi:10.1109/LSP.2019.2918933

Explainable AI (XAI): Personalized and Transparent Models

Building Trust Through Personalized Explainability

For AI systems to gain widespread acceptance, their predictions must be explainable. We develop personalized explanations tailored to individual users. By measuring the reduction in conditional entropy, this method ensures explanations align with the user’s knowledge and needs.

Key Features:

  • User-Centric Explanations: Accounts for user background and preferences.
  • Model-Agnostic Framework: Compatible with any machine learning model.
  • Efficient Implementation: Requires minimal training data—just data points, predictions, and user-provided summaries.

  • “Explainable Empirical Risk Minimization”
    L. Zhang, G. Karakasidis, A. Odnoblyudova, et al., Neural Computing and Applications, 2024.
    doi:10.1007/s00521-023-09269-3

  • “An Information-Theoretic Approach to Personalized Explainable Machine Learning”
    A. Jung and P. H. J. Nardelli, IEEE Signal Processing Letters, 2020.
    doi:10.1109/LSP.2020.2993176