Split Learning Explainer
Split learning divides a neural network at a chosen cut layer. A client runs the early layers on local data, sends the resulting activation to a server, and the server completes the forward pass and part of backpropagation.
The server returns a gradient at the cut so the client can update its local layers. Raw input records do not have to leave the client, and each side holds only part of the model during the protocol.
This design can help resource-limited clients collaborate with stronger servers. The cut location affects client workload, communication volume, model exposure, latency, and how much information the intermediate activation may reveal.
Activations and gradients are not automatically private; reconstruction and inference attacks may still apply. Split learning therefore needs a threat model, secure transport, access controls, leakage testing, and careful handling of labels and identifiers.
Split learning divides a neural network between participants so raw data stays with the client while intermediate activations cross the boundary.
Key terms include cut layer, client, server, activation, gradient, forward pass, backpropagation, communication and leakage.
The client computes early layers and sends an intermediate activation.
The server finishes the model and returns a gradient for client-side updating.
Keeping raw data local does not make activations or gradients harmless.
Leakage tests and transport security remain necessary.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.