Federated Meta-Learning

The foundational algorithm that Federated Meta-Learning adapts to decentralized settings. MAML finds an optimal model initialization such that a few gradient steps on a new task produce a high-performing specialized model. In healthcare, this enables a model trained across hospitals to rapidly adapt to a new clinic's patient demographics with minimal local data.

• Inner loop: Task-specific adaptation via few-shot learning
• Outer loop: Meta-optimization across tasks to improve initialization
• Key metric: Post-adaptation accuracy after 1-5 gradient steps

A direct marriage of Federated Averaging and MAML that explicitly optimizes for personalization. Unlike standard FedAvg which minimizes global loss, Per-FedAvg finds a shared initial model that each client can fine-tune locally with minimal computational overhead. This is critical for hospitals with limited GPU resources that still require site-specific model adaptation.

• Objective: Minimize local loss after one step of gradient descent
• Advantage over FedAvg: Handles statistical heterogeneity natively
• Use case: Rare disease detection where each hospital sees different patient subpopulations

Complements meta-learning by addressing feature space misalignment across institutions. While meta-learning focuses on rapid adaptation, FTL handles scenarios where different hospitals collect entirely different feature sets. A model trained on genomic data at one site can transfer knowledge to a site with only imaging data through common representation learning.

• Feature-based FTL: Aligns heterogeneous feature spaces via domain adaptation
• Parameter-based FTL: Shares pre-trained layers across related tasks
• Clinical example: Transferring knowledge from radiology models to pathology models

The evaluation paradigm that meta-learning optimizes for. In clinical settings, rare conditions may have only 5-10 annotated examples across an entire hospital network. Few-shot learning measures a model's ability to generalize from these minimal examples. Federated meta-learning is specifically designed to produce models that excel in this low-data regime.

• N-way K-shot: Classify N classes with K examples each
• Prototypical networks: Learn a metric space where classification is distance-based
• Clinical relevance: Diagnosing rare genetic disorders from limited case histories

An alternative personalization architecture where a central hypernetwork generates the weights of client-specific models. Instead of sharing model parameters, clients share client descriptors (e.g., demographic statistics, data distribution summaries). The hypernetwork learns to map these descriptors to personalized model weights, enabling structured adaptation without iterative local fine-tuning.

• Privacy benefit: Raw model weights never transmitted
• Structured personalization: Explicit conditioning on known client characteristics
• Trade-off: Requires meaningful client descriptors to be effective

Treats each client as a separate but related task, learning shared representations while allowing task-specific parameters. Unlike meta-learning which optimizes for adaptation speed, multi-task learning optimizes for simultaneous performance across all clients. The two approaches can be combined: meta-learning provides the initialization, and multi-task structure maintains client-specific heads.

• Hard parameter sharing: Shared feature extractor, task-specific classifiers
• Soft parameter sharing: Regularized distance between client models
• Clinical application: Shared organ segmentation model with hospital-specific refinements