From 2 to 7 December 2025, SmartCHANGE was represented by the Università della Svizzera italiana (USI) at the 39th Annual Conference on Neural Information Processing Systems (NeurIPS) in San Diego, USA.
About the event
The Conference on Neural Information Processing Systems (NeurIPS) is the world’s premier interdisciplinary meeting for machine learning (ML) and computational neuroscience. Founded in 1987, it has evolved into a massive multi-track event featuring invited talks, symposia, and presentations of peer-reviewed research.
The 2025 edition brought together global experts across fields such as deep learning, optimisation, computer vision, and the life sciences. Beyond the core academic tracks, the conference featured a professional exposition and topical workshops, providing a key space for both theoretical breakthroughs and practical applications of machine learning.
SmartCHANGE at the event
Representing the SmartCHANGE project, researchers from the Università della Svizzera italiana (USI), Dario Fenoglio, Martin Gjoreski, and Marc Langheinrich, attended the conference to present a significant advancement in distributed AI. On 5 December, the team presented their publication in the Exhibit Hall: "FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts".
The presentation addressed a critical hurdle in Federated Learning (FL): the drop in model accuracy when data across different clients is not identical (non-IID). The SmartCHANGE team introduced FLUX, a novel framework that uses privacy-preserving clustering to handle distribution shifts without requiring prior knowledge of the data types or the number of clusters.
Key Achievements of FLUX:
- Superior Accuracy: Achieved gains of up to 23 percentage points over state-of-the-art baselines.
- Adaptability: Supports "test-time adaptation," allowing new, unlabeled clients to benefit from the system immediately.
- Efficiency: Maintains a communication and computational footprint comparable to standard methods like FedAvg.
Find the full paper and technical details at this link.
By engaging with machine learning practitioners and industry experts, the SmartCHANGE team fostered new collaborations and gained high-level visibility for their work in making AI more robust and scalable for real-world health and science applications.
Learn more on the official event website