Horizontal Federated Learning

A federated learning paradigm where collaborating parties hold data with different features about the same set of overlapping entities. Unlike horizontal FL, vertical FL requires entity alignment—a privacy-preserving step to match common users across datasets without revealing identities. Training typically employs a split neural network architecture, where each party maintains its own bottom model and exchanges intermediate activations rather than raw data. This is common in cross-industry collaborations, such as a bank and a retailer jointly building a fraud model using complementary financial and purchasing features.

A technique that applies transfer learning within a federated framework to handle scenarios where clients have datasets with different feature spaces and different label spaces. FTL leverages a common representation learned from a related domain to bridge the gap between heterogeneous parties. The process involves:

• Learning a common feature subspace using overlapping samples or auxiliary data
• Transferring knowledge from source domains with rich labels to target domains with scarce labels
• Enabling collaboration even when no single party has both comprehensive features and sufficient labels This is particularly valuable in rare disease research, where individual institutions may lack sufficient labeled cases to train a model independently.

A federated learning topology designed for a small number of reliable institutional clients—typically 2 to 100—such as hospitals, pharmaceutical companies, or financial institutions. Key characteristics include:

• Persistent participation: All clients are available for every training round
• Large, curated datasets: Each silo holds substantial, well-maintained data
• High compute resources: Clients possess server-grade infrastructure
• Centralized orchestration: A trusted server coordinates the aggregation Cross-silo FL is the dominant paradigm in healthcare biomarker discovery, where a handful of hospital systems collaboratively train diagnostic models without exposing patient records.

A federated learning topology designed for massive scale—potentially millions of unreliable edge devices such as smartphones, IoT sensors, or wearables. This paradigm must contend with:

• Intermittent availability: Devices participate only when idle, charging, and on Wi-Fi
• Limited compute: Training occurs on mobile processors with constrained memory
• High dropout rates: A significant fraction of clients never complete a round
• Communication efficiency: Model updates must be compressed to minimize bandwidth Google's Gboard keyboard uses cross-device FL to improve next-word prediction across billions of devices without uploading typed text to the cloud.

A federated learning strategy that moves beyond a single global model to create tailored local models for individual clients. In horizontal FL, a one-size-fits-all global model may perform poorly for clients with highly divergent data distributions. pFL addresses this through:

• Model interpolation: Blending a global model with a local-only model
• Meta-learning: Training an initialization that adapts quickly to each client with few gradient steps
• Multi-task learning: Treating each client as a separate task with shared representations
• Clustered FL: Grouping similar clients and training separate models per cluster In biomarker identification, pFL allows each hospital to maintain a model optimized for its specific patient demographics while still benefiting from broader collaborative training.

A privacy-preserving distributed deep learning technique where a neural network is partitioned between a client and server. The client processes data through the initial layers and sends only intermediate activations (smashed data) to the server, which completes the forward pass and backpropagation. Key properties:

• No raw data sharing: The server never sees the original inputs
• No model sharing: Clients never receive the full model architecture
• Sequential training: Unlike FL, clients train sequentially rather than in parallel
• Communication overhead: Only activations and gradients of the cut layer are exchanged Split learning is particularly useful when clients have limited compute resources and cannot train a full model locally, making it complementary to horizontal FL in resource-constrained healthcare settings.