split learning

**Split Learning** is a **distributed learning technique that splits a neural network at a "cut layer" between the client and server** — the client processes data through the lower layers (keeping data private), sends intermediate activations (smashed data) to the server, which completes the forward pass and backpropagation. **How Split Learning Works** - **Client Side**: Forward pass through layers 1-$k$ on private data → produce activations $a_k$. - **Server Side**: Receive $a_k$, forward through layers $k+1$-$L$, compute loss, backpropagate to layer $k+1$. - **Gradient Return**: Server sends $ abla a_k$ (gradient of loss w.r.t. activations) back to client. - **Client Backward**: Client backpropagates through layers 1-$k$ using $ abla a_k$. **Why It Matters** - **Low Client Compute**: Client only runs part of the model — suitable for resource-constrained edge devices. - **Privacy**: Raw data never leaves the client — only intermediate activations are shared. - **Caveat**: Intermediate activations can leak information — additional protections (noise, quantization) may be needed. **Split Learning** is **dividing the neural network** — clients process their data through the bottom half, servers complete the computation through the top half.