Device-Edge-Cloud Continuum

The strategic division of a deep neural network's computational graph into distinct segments for distributed execution. A model is sliced at a specific layer, with the head (early layers) executing on the device and the tail (later layers) on an edge server. This minimizes data transmitted over the network compared to raw sensor data offloading.

• Horizontal Partitioning: Splitting by layer depth (e.g., after a bottleneck layer)
• Vertical Partitioning: Distributing parallel branches of a model across nodes
• Key metric: minimizing transmission tensor size while preserving accuracy

A specific model partitioning technique that slices a deep neural network at a designated bottleneck layer. The device executes up to the split point, then transmits the compressed intermediate feature map to an edge server for the remaining computation. This is the foundational mechanism of split computing.

• Reduces transmitted data by 10x-100x compared to raw input offloading
• The split point is chosen to balance compute load and bandwidth consumption
• Commonly used with lightweight encoders on-device and heavy decoders server-side

An adaptive decision-making process that determines in real-time whether to execute an inference task locally or offload it to an edge server. The decision engine monitors network telemetry (latency, bandwidth), device load (battery, thermal state), and model characteristics to optimize for a target latency budget.

• Uses heuristic thresholds or lightweight ML classifiers for sub-millisecond decisions
• Critical for mobile AR/VR workloads where network conditions fluctuate rapidly
• Often paired with channel-aware compression to adapt to link quality

A Multi-access Edge Computing server deployed at the network edge, providing cloud-computing capabilities within the radio access network. MEC servers sit between the device and the centralized cloud, offering ultra-low latency (often < 10ms) and high-bandwidth connectivity for inference offloading.

• Standardized by ETSI, deployed at base stations or aggregation points
• Enables collaborative inference for latency-critical applications like autonomous driving
• Provides an IT service environment for running containerized inference workloads

Techniques applied to the activations transmitted at the partition point in split computing to reduce bandwidth consumption. After the device executes the model head, the output tensor is compressed before transmission and decompressed at the edge server before continuing inference.

• Quantization: Reducing activation precision (e.g., FP32 to INT8) before transmission
• Entropy coding: Lossless compression using statistical redundancy (e.g., Huffman, arithmetic coding)
• Dimensionality reduction: Applying PCA or learned projections to shrink the feature vector
• Trade-off: compression ratio vs. downstream accuracy degradation

A model slicing strategy that considers Quality of Service requirements—such as latency, accuracy, and energy constraints—to dynamically select the optimal partition point for each inference request. Unlike static partitioning, this adapts to per-request SLAs.

• A video analytics app may prioritize accuracy for security alerts but latency for real-time previews
• The partition point shifts deeper (more on-device) when network conditions degrade
• Often implemented via a lookup table or lightweight neural network trained on profiling data