Multi-Task Federated Learning

MTFL architectures typically learn a global feature extractor shared across all clients and tasks, while maintaining task-specific heads for individual predictions. This forces the network to discover a common latent space that captures the underlying structure of related clinical problems—such as learning a universal embedding for chest X-rays that supports both nodule detection and pleural effusion classification simultaneously. The shared layers benefit from aggregated data across all sites, while specialized layers adapt to local label distributions.

Unlike independent single-task models, MTFL explicitly encodes the statistical relationships between tasks to enable knowledge transfer. Common strategies include:

• Hard parameter sharing: A single backbone with multiple output branches
• Soft parameter sharing: Separate networks with regularization penalties that encourage weight similarity
• Task covariance matrices: Learning a probabilistic prior over task relationships In clinical settings, this allows a model trained for mortality prediction to improve length-of-stay forecasting by exploiting their correlation.

MTFL is designed for scenarios where not all clients possess labels for all tasks. A hospital may have annotated data for diabetic retinopathy but not for glaucoma detection, while another institution has the inverse. The framework handles this partial label coverage by:

• Masking loss terms for missing tasks during local training
• Aggregating only the parameters relevant to tasks present at each client
• Using task-level participation weighting in the global aggregation step This mirrors real-world clinical data silos where annotation resources are unevenly distributed.

MTFL enables indirect knowledge transfer between clients that never share raw data. A client with abundant labels for Task A contributes to the shared representation, which in turn improves performance on Task B at a different client with limited Task B labels. This is achieved through the federated aggregation of shared parameters, where gradient updates from one task propagate benefits to others via the common feature space. The mechanism is particularly valuable for rare disease diagnosis, where no single institution has sufficient cases to train a robust model independently.

MTFL naturally supports client-level personalization by allowing each institution to assign different loss weights to tasks based on local clinical priorities. A research hospital may emphasize tumor segmentation accuracy, while a community clinic prioritizes triage classification recall. The global aggregation reconciles these preferences by:

• Computing a Pareto-optimal solution across competing task objectives
• Using gradient normalization to prevent any single task from dominating shared parameters
• Allowing clients to maintain private task-weighting vectors that never leave their infrastructure

MTFL reduces communication overhead compared to training separate federated models for each task. By co-training related tasks within a single architecture, the framework transmits one set of shared parameters instead of N independent models. Advanced implementations use task grouping algorithms that cluster correlated tasks together and assign uncorrelated tasks to separate model branches, optimizing the accuracy-communication trade-off. This is critical in healthcare networks where bandwidth constraints and HIPAA-compliant transmission protocols impose strict limits on data exchange frequency.