Federated Continual Learning

The central challenge in FCL is preventing a global model's performance on Task A from degrading sharply when it learns Task B from a new stream of client data. This is exacerbated by the federated setting, where the server cannot replay historical patient data. Solutions involve regularization-based approaches like Elastic Weight Consolidation (EWC) applied to the global model, penalizing changes to parameters critical for previous tasks. Alternatively, dynamic architecture methods expand the model capacity for new tasks while freezing old weights, and rehearsal-based methods use a small, privacy-compliant proxy dataset or generative replay to interleave past knowledge during training.

FCL deals with a dual heterogeneity problem: spatial non-IIDness (data distribution skew across clients at a given time) and temporal non-IIDness (concept drift over time). A model must adapt to evolving clinical practices, new diagnostic criteria, or novel disease outbreaks without forgetting rare historical conditions. This requires the aggregation server to distinguish between a genuine temporal shift requiring global model adaptation and a transient local distribution anomaly at a single hospital.

Since raw data cannot be centralized for experience replay, FCL relies on pseudo-rehearsal techniques. Clients or the server can train a generative model (like a GAN or VAE) to synthesize representative samples from previous tasks. These synthetic samples are then interleaved with current task data during local training to reinforce old decision boundaries. Alternatively, federated knowledge distillation uses the previous global model as a teacher, enforcing consistency between its soft predictions and the current model's outputs on new data to preserve past knowledge.

In real-world healthcare deployments, new tasks often arrive asynchronously at different clients. One hospital may introduce a new classification task while others continue with existing ones. This task-incremental or class-incremental scenario requires the global model to seamlessly integrate new output heads or classification layers without requiring all clients to participate in every update round. The aggregation protocol must handle heterogeneous model architectures where some clients have task-specific parameters that others lack.

A class of FCL algorithms adapts Gradient Episodic Memory (GEM) to the federated context. The server maintains a memory buffer of parameter gradients from previous tasks. During local training on a new task, the optimization is constrained to ensure that the loss on previous tasks does not increase. This is formulated as a quadratic programming problem: find a gradient update that minimizes the new task loss while projecting onto the feasible region defined by negative dot products with all previous task gradients.

Building on Synaptic Intelligence (SI) , this approach computes a per-parameter importance weight that reflects each synapse's contribution to past task performance. During training on a new task, a quadratic penalty term discourages significant changes to high-importance parameters. In the federated setting, these importance weights can be aggregated alongside model weights using Federated Averaging, creating a global importance mask that protects consolidated knowledge across the entire network of hospitals.

The primary adversary of continual learning. When a neural network is trained sequentially on new tasks, its performance on previously learned tasks degrades rapidly as weights are overwritten.

• Stability-Plasticity Dilemma: The core tension—how to learn new information without erasing old knowledge.
• Example: A diagnostic model trained on chest X-rays, then fine-tuned on knee MRIs, may lose its ability to detect pneumonia entirely.
• Severity: Forgetting can be near-total; accuracy on old tasks can drop from 95% to chance levels after only a few sequential training rounds.

A rehearsal-based strategy that stores a small memory buffer of representative samples from previous tasks and interleaves them during training on new tasks.

• Federated Constraint: Storing raw patient data violates privacy principles. Federated systems must use pseudo-rehearsal—generating synthetic samples from a generative model instead of storing real data.
• Generative Replay: A generative adversarial network or variational autoencoder is trained to produce realistic samples mimicking prior task distributions without retaining actual patient records.
• Trade-off: Memory buffer size directly impacts retention quality; larger buffers improve recall but increase storage and privacy risk.

A temporal shift in the statistical relationship between input features and target labels over time, distinct from the catastrophic forgetting problem.

• Real vs. Virtual Drift: Real drift changes P(Y|X)—the actual diagnostic criteria evolve. Virtual drift changes P(X) only—patient demographics shift but disease definitions remain constant.
• Clinical Example: During a pandemic, the prior probability of respiratory conditions spikes (virtual drift), while updated testing protocols may redefine what constitutes a positive case (real drift).
• FCL Integration: Continual learning systems must distinguish between drift requiring adaptation and noise triggering unnecessary retraining.

A communication paradigm where clients share soft label predictions on a public unlabeled dataset instead of model parameters, enabling heterogeneous continual learning.

• Mechanism: Each client's model acts as a teacher, generating class probability distributions for a shared public dataset. The server aggregates these soft labels to train a global student model.
• Continual Advantage: Distillation naturally preserves knowledge across tasks because the student learns from the aggregated output distribution, which encodes inter-class relationships from all prior tasks.
• Privacy Benefit: No raw data or model gradients are exchanged; only prediction vectors leave the client, reducing the attack surface for model inversion.