In the current landscape of distributed artificial intelligence, one of the most complex challenges is enabling multiple clients, each with isolated and private data, to collaborate in training models that must continuously adapt to new tasks. This scenario, known as Federated Class Incremental Learning (FCIL), combines the difficulties of Federated Learning (FL) with those of Continual Learning (CL). However, their convergence introduces two coupled sources of interference: spatial interference, caused by client heterogeneity, and temporal interference, arising from the sequential arrival of tasks. Together, they lead to Spatial-Temporal Catastrophic Forgetting (ST-CF).
Existing approaches typically address these two problems separately, implementing client-side mechanisms that increase computational or communication overhead. Moreover, they rarely regulate the directional interactions among updates during aggregation. To overcome this limitation, SUM (Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors) emerges as a purely server-side framework that reinterprets FCIL as a unified multi-task learning problem. In this view, both client updates and task updates are represented as adaptation vectors in a shared parameter space. SUM applies geometric surgery on these vectors during aggregation: Spatial SUM mitigates client-level interference within each round, while Causal Online Temporal SUM removes cross-task interference over time, without requiring additional client-side computation, communication, or memory beyond standard federated training.
From a technical standpoint, SUM operates directly on the directions of the adaptation vectors, adjusting their orientations to minimize conflicts. This yields empirical improvements of up to 22% over prior FCIL methods on vision and language benchmarks, even in scenarios with unreliable or dynamic clients. Computational efficiency is maintained, as all additional processing is concentrated on the server, facilitating integration into existing cloud infrastructures.
For companies developing distributed intelligent systems, SUM opens new possibilities. At Q2BSTUDIO, as a company specialized in custom software development, we understand that solutions like SUM can be embedded into platforms requiring continuous learning without compromising data privacy. For example, in environments where multiple branches or IoT devices collaborate to improve an anomaly detection model, geometric surgery prevents the knowledge from one branch interfering with another, while temporal adaptation allows incorporating new anomaly classes without forgetting previous ones.
Implementing SUM greatly benefits from a robust cloud infrastructure. The ability to scale aggregation servers, manage asynchronous communication, and ensure availability of updated models is key. Therefore, the cloud AWS and Azure services we offer at Q2BSTUDIO are ideal for deploying such federated architectures. Moreover, data security is paramount: in a federated environment, each client retains data locally, but vector aggregation must be protected against inference or poisoning attacks. Our cybersecurity solutions help secure the process, from client authentication to communication encryption.
Another relevant dimension is monitoring and analyzing model performance over time. BI and Power BI tools enable visualization of metrics such as forgetting rate, per-task accuracy, or client contribution, facilitating informed decisions on when to retrain or redistribute load. It is even possible to integrate AI agents that, based on analysis of these metrics, dynamically adjust the parameters of geometric surgery, optimizing the balance between stability and plasticity.
At Q2BSTUDIO, we design custom artificial intelligence solutions for companies that need federated and incremental learning systems. Incorporating techniques like SUM into our developments allows our clients to benefit from models that evolve without forgetting, collaborate without interfering, and adapt without saturating edge device resources. Whether in industrial vision applications, multilingual natural language processing, or distributed recommendation systems, unified geometric surgery represents a significant advance toward more efficient, scalable, and privacy-respecting artificial intelligence.
In summary, SUM offers an elegant and practical solution to the problem of spatial-temporal catastrophic forgetting in federated class incremental learning. By concentrating all interference mitigation logic on the server, it simplifies deployment on heterogeneous clients and reduces operational costs. For organizations looking to implement AI systems that learn continuously and collaboratively, combining SUM with a suitable cloud infrastructure, robust cybersecurity measures, and BI tools provides a solid roadmap. At Q2BSTUDIO, we are ready to support that journey with our technical capabilities and experience in digital transformation projects.
