Progressive Feature Transmission for Split Classification at the Wireless Edge

Abstract

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

MLRS-RL: An Energy-Efficient Multilevel Routing Strategy Based on Reinforcement Learning in Multimodal UWSNs

“…The split inference is an advanced edge AI architecture designed to optimize energy consumption in edge devices [5], [18]- [20]. Specifically, the overall model is partitioned into two components, with one part being processed by the edge device and the other by the server.…”

“…In this case, the classification model should be divided into two parts to implement split inference, i.e., device sub-model and server sub-model. Following the widely used designs [5], [18], the splitting point is set after the CONV layers, where the output of the last CONV layer is a feature map with height L h and width L w . Each element in the feature map is quantized with a sufficiently high resolution of Q bits such that quantization errors are negligible.…”

An Integrated Method for River Water Level Recognition from Surveillance Images Using Convolution Neural Networks

“…2) Discrimination Gain: The sensing accuracy is largely determined by the discernibility between a pair of classes that can be measured by discrimination gains computed as their symmetric Kullback-Leibler (KL) divergence [38]. Considering sensor k, the local discrimination gain between classes ℓ and ℓ ′ , denoted as G k (ℓ, ℓ ′ ), can be computed as…”

“…Rayleigh fading, representing spatial diversity from rich scattering and spatially separated sensors. Last, the E2E sensing accuracy is measured by the popular metric of sensing uncertainty that is computed as the entropy of posteriors of object classes conditioned on observations [38], [39].…”

“…The distribution of f k is given as follows. First, let g ∈ R M be the ground-truth feature map corresponding to the current object and be assumed to have a uniform prior distribution over L classes [24], [38]:…”