Model Splitting

The overarching distributed training paradigm where a model is divided across client and server. Unlike federated learning, which shares model updates, split learning shares only smashed data (intermediate activations) and gradients at the cut layer. This prevents raw data exposure and reduces client-side compute load, making it ideal for resource-constrained medical devices.

The critical architectural decision of where to partition the network. An early cut (e.g., after the first convolutional block) minimizes on-device compute but transmits larger feature maps, increasing bandwidth. A deep cut (e.g., before the classification head) reduces transmission size but demands more device processing power. Selection is a direct trade-off between latency, privacy, and energy consumption.

The intermediate activation tensor produced by the final client-side layer before transmission to the server. This data is a highly abstract, non-human-readable representation of the input. While it does not contain raw pixels or text, model inversion attacks can potentially reconstruct input features from smashed data, necessitating additional privacy measures like differential privacy.

A specialized splitting strategy for encoder-decoder architectures common in medical image segmentation. The encoder is placed on the edge device to extract features locally, and the decoder resides on a powerful server. This allows high-resolution diagnostic outputs without transmitting massive raw scans, balancing HIPAA compliance with the need for complex generative inference.

An execution model that distributes an AI workload across different processors on a System-on-Chip (SoC). In a split model, the client-side layers might run on a low-power NPU or DSP, while the server-side layers leverage high-throughput GPUs. Efficient delegation is critical to meeting strict latency budgets for real-time medical applications.