Semantic Split Computing

The core innovation of semantic split computing is transmitting only the latent feature tensor from an intermediate neural network layer, not the raw input. This bottleneck representation is typically 100x to 1000x smaller than the source data. For example, a 224x224 RGB image (150k values) can be compressed to a 7x7x512 feature map (25k values) while preserving task-relevant semantics. The split layer is chosen to balance compression ratio against downstream task accuracy.

By transmitting abstract feature vectors instead of raw sensor data, semantic split computing provides an inherent privacy barrier. The network operator never receives the original image, audio, or RF waveform. Key properties include:

• Input obfuscation: Raw data stays on-device
• Non-invertibility: Intermediate features cannot be trivially reconstructed into the original input
• Differential privacy guarantees: Noise can be injected at the split layer to provide formal privacy bounds This makes the architecture suitable for GDPR-compliant and HIPAA-compliant deployments.

Semantic split computing enables dynamic workload partitioning between resource-constrained edge devices and powerful cloud servers. The edge executes the initial feature extraction layers (often quantized for efficiency), while the cloud handles the computationally intensive semantic reasoning layers. This is critical for:

• TinyML devices: Microcontrollers running only 2-3 initial conv layers
• Mobile AR/VR: Offloading heavy transformer decoders to edge servers
• Autonomous systems: Splitting perception models between onboard NPU and remote inference engines

Unlike traditional systems that separately compress then channel-code data, semantic split computing enables joint optimization of the feature extractor and the wireless transmitter. The split layer's output can be directly mapped to channel symbols using learned constellation designs. This end-to-end training approach maximizes task performance under specific channel conditions (SNR, fading, interference) rather than optimizing for abstract bit-error metrics. Research shows 10-15 dB gain in effective SNR for image classification tasks compared to separate source-channel coding.

The split layer position is not fixed but can be dynamically selected based on:

• Channel quality: Deeper splits (more compression) for poor channel conditions
• Task complexity: Earlier splits for simple classification, later splits for detailed reconstruction
• Available edge compute: More layers on-device when local GPU/NPU capacity permits
• Privacy requirements: Earlier splits when stricter input privacy is needed This adaptability is implemented through multi-exit architectures where the model has multiple potential split points, each trained for different operating regimes.

To maximize transmission efficiency, semantic split computing employs learned entropy models that estimate the probability distribution of feature tensors. This enables arithmetic coding of the bottleneck representation, achieving near-optimal compression rates. Advanced implementations use:

• Hyperprior networks: A secondary autoencoder that transmits side information about feature statistics
• Context-adaptive coding: Exploiting spatial correlations within feature maps
• Variable-rate training: A single model that can operate across a range of bitrates by conditioning on a rate parameter

A neural network component that extracts and compresses the essential meaning from a source signal, discarding task-irrelevant information before transmission. In a split computing architecture, the encoder is typically partitioned: a lightweight on-device encoder extracts initial features, while a more powerful server-side encoder refines them. This bottleneck representation dramatically reduces bandwidth compared to raw data transmission.

A deep learning paradigm that replaces separate source and channel coding blocks with a single neural autoencoder, directly mapping source data to channel symbols. In semantic split computing, JSCC enables the intermediate feature vector to be transmitted as channel-optimized symbols rather than bits, providing graceful degradation under poor channel conditions instead of the cliff effect seen in traditional digital communications.

An information-theoretic framework that learns a compressed, stochastic latent representation maximally predictive of a target task while discarding irrelevant data. In split computing, VIB provides the mathematical foundation for determining the optimal split point in a deep model, balancing the trade-off between the compression rate of transmitted features and the accuracy of the downstream task.

A transmission paradigm where information is encoded and decoded based on its effectiveness in achieving a specific receiver task, rather than on symbol-level accuracy. Semantic split computing operationalizes this concept by transmitting only the intermediate features relevant to the server-side task. Key characteristics include:

• Task-specific compression rather than universal reconstruction
• Resilience to semantic noise that does not affect task outcomes
• Drastic bandwidth reduction for inference-heavy edge applications

A shared, structured repository of background knowledge, ontologies, and common sense used by both transmitter and receiver to interpret transmitted meaning. In split architectures, the SKB enables the edge device and server to maintain contextual alignment without transmitting redundant information. The edge encoder can transmit compact indices or references to shared concepts rather than full feature vectors, further reducing communication overhead.

Compression techniques including quantization-aware training, pruning, and knowledge distillation specifically for deploying neural components on resource-constrained edge hardware. For semantic split computing, these methods are critical for the on-device encoder segment, which must operate within strict power and memory budgets. Techniques include:

• INT8 quantization of encoder weights
• Structured pruning of redundant feature extractors
• Distillation from larger teacher encoders into compact student models