SUM: серверный метод против «забывания» в федеративном непрерывном обучении, +22%

A team of researchers published a paper on arXiv in July 2026 describing SUM (Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors) — a server-side framework for federated continual learning that reduces "catastrophic forgetting" and delivers a quality gain of up to 22% over previous FCIL methods.

What the SUM method solves

The SUM method tackles "Spatial-Temporal Catastrophic Forgetting" (ST-CF) — the loss of knowledge at the intersection of two learning approaches. Federated Class Incremental Learning (FCIL) combines federated learning (FL), where a model trains on isolated data from different clients, with continual learning (CL), where tasks arrive sequentially.

According to the arXiv paper, this combination produces two sources of interference at once. The first is spatial interference from heterogeneous clients with dissimilar data. The second is temporal interference from sequential tasks, where new knowledge displaces old knowledge. Previous methods addressed these problems with different mechanisms, often burdening clients with extra computation or data exchange, and did not regulate directional interactions between updates during aggregation.

How "geometric surgery" works

SUM runs entirely on the server side and corrects "adaptation vectors" during update aggregation. The authors propose treating FCIL as a single multi-task learning problem, where updates from both clients and tasks are vectors in a shared parameter space.

• SUM is a server-side framework, with no extra computation, communication, or memory on the client
• Spatial SUM removes inter-client interference within each training round
• Temporal SUM (causal online) eliminates inter-task interference over time
• Quality gain — up to 22% over previous FCIL methods
• Tests — on computer vision and language benchmarks, with robustness to unreliable clients

How much better is this than previous approaches

SUM outperforms previous FCIL methods by up to 22% across a range of vision and language benchmarks — while preserving computational efficiency. The key difference: all the "repair" happens during aggregation on the server, so clients don't need to do extra computation or send additional data.

"Spatial SUM reduces interference at the client level within each round, while causal online

Temporal SUM removes inter-task interference over time without additional computation, communication, or client memory beyond standard federated learning," the arXiv paper's abstract states.

Separately, the authors note the method's robustness to unreliable clients — those that produce noisy or incomplete updates. This matters for real-world distributed systems, where some devices are always operating unstably.

What this means

Federated continual learning is a scenario for systems that need both to learn from scattered private data and to adapt to new tasks without retraining from scratch. SUM shows that forgetting can be countered on the server, without shifting the load onto client devices — making the approach applicable to phones, IoT, and other resource-constrained devices.

Frequently asked questions

What is Federated Class Incremental Learning?

FCIL is a combination of federated learning (a model trains on isolated data from many clients without collecting it in one place) and class-incremental learning (new classes and tasks arrive over time). The main challenge is "catastrophic forgetting," when a model loses old knowledge while learning new tasks.

How does SUM differ from previous FCIL methods?

SUM operates only on the server and combines the fight against spatial and temporal interference into a single mechanism — "geometric surgery" over adaptation vectors. Previous approaches used separate mechanisms and often burdened clients, whereas SUM requires no additional computation, communication, or memory from them.