Progressive Feature Transmission for Split Classification at the Wireless Edge

We consider the scenario of inference at wireless edge, in which devices are connected to an edge server and ask carry out remote classification, that is, classify data samples available devices. This requires upload high-dimensional features over resource-constrained channels, creates a communication bottleneck. The conventional feature pruning solution would require device have access model, is not current split scenario. To address this issue, we propose progressive transmission (ProgressFTX) protocol, minimizes overhead by progressively transmitting until target confidence level reached. A control policy proposed accelerate inference, comprising two key operations: importance-aware selection transmission-termination control. For former, it shown selecting most important features, characterized largest discriminant gains corresponding dimensions, achieves sub-optimal performance. latter, exhibit threshold structure. Specifically, stopped when incremental uncertainty reduction further outweighed its cost. indices selected decision fed back each slot. first derived for tractable case linear then extended more complex classification using convolutional neural network. Both Gaussian fading channels considered. Experimental results obtained both statistical model real dataset. It ProgressFTX can substantially reduce latency compared random strategies.