Source-linked AI summary
Task-Oriented Sensing, Computation, and Communication Integration for Multi-Device Edge AI
Dingzhu Wen, Peixi Liu, Guangxu Zhu, Yuanming Shi, Jie Xu, Yonina C. Eldar, Shuguang Cui
TL;DR
The paper addresses low-latency edge AI inference when sensing, computation, and communication are coupled in multi-device split inference. It models multi-view ISAC feature acquisition and offloading with discriminant gain, then optimally allocates power, time, and quantization bits using a sum-of-ratios method. Experiments on human motion recognition verify the proposed optimal ISCC scheme.
Problem
Task-oriented ISCC for edge inference remains insufficiently developed because prior work overlooks computation and focuses mainly on offline training.
Method
The paper uses discriminant gain to model classification performance and jointly optimizes sensing and communication power, time, and quantization-bit allocation through the sum-of-ratios method.
Results
The proposed optimal ISCC scheme is verified through extensive human motion recognition experiments.
Takeaways & Limitations
Jointly allocating sensing, communication, and quantization resources enhances real-time inference accuracy measured by discriminant gain.
Abstract
from arXiv · showhide
This paper studies a new multi-device edge artificial-intelligent (AI) system, which jointly exploits the AI model split inference and integrated sensing and communication (ISAC) to enable low-latency intelligent services at the network edge. In this system, multiple ISAC devices perform radar sensing to obtain multi-view data, and then offload the quantized version of extracted features to a centralized edge server, which conducts model inference based on the cascaded feature vectors. Under this setup and by considering classification tasks, we measure the inference accuracy by adopting an approximate but tractable metric, namely discriminant gain, which is defined as the distance of two classes in the Euclidean feature space under normalized covariance. To maximize the discriminant gain, we first quantify the influence of the sensing, computation, and communication processes on it with a derived closed-form expression. Then, an end-to-end task-oriented resource management approach is developed by integrating the three processes into a joint design. This integrated sensing, computation, and communication (ISCC) design approach, however, leads to a challenging non-convex optimization problem, due to the complicated form of discriminant gain and the device heterogeneity in terms of channel gain, quantization level, and generated feature subsets. Remarkably, the considered non-convex problem can be optimally solved based on the sum-of-ratios method. This gives the optimal ISCC scheme, that jointly determines the transmit power and time allocation at multiple devices for sensing and communication, as well as their quantization bits allocation for computation distortion control. By using human motions recognition as a concrete AI inference task, extensive experiments are conducted to verify the performance of our derived optimal ISCC scheme.
I. INTRODUCTION
The paper motivates task-oriented ISCC for edge inference by jointly optimizing sensing, computation, and communication under accuracy and latency requirements. It introduces a multi-view ISAC edge-inference system and uses discriminant gain to guide an optimal resource-allocation design.
- Motivation: Existing split-inference designs optimize computation or communication separately, although sensing, computation, and communication jointly determine feature distortion and inference performance.The three processes are coupled through sensing noise, quantization distortion, communication capacity, and delay constraints.
- Motivation: Task-oriented ISCC targets inference accuracy and latency rather than throughput, jointly designing sensing, quantization, and transmission under device-resource and low-latency constraints.In edge AI, sensing and communication share the goal of improving inference performance.
- Related work: Prior task-oriented ISAC studies overlook computation and largely focus on offline training, leaving task-oriented ISCC for edge inference insufficiently explored.The paper identifies this gap as motivation for its edge-inference design.
- System model: The proposed system uses multiple ISAC devices for multi-view radar sensing, quantized feature offloading, and centralized inference on cascaded feature vectors.The objective is real-time classification under a latency constraint.
- Accuracy metric: Discriminant gain measures class discernibility in Euclidean feature space and is adopted as an approximate but tractable proxy for inference accuracy.The paper derives a closed-form expression capturing the effects of sensing noise, quantization distortion, and communication capacity.
- Contributions: The paper formulates a non-convex ISCC design and applies the sum-of-ratios method to jointly allocate sensing and communication resources and quantization bits.The resulting scheme addresses transmit power, time allocation, and computation-distortion control.
- Validation: Experiments report that maximizing discriminant gain is effective for SVM and MLP inference accuracy, while the optimal ISCC scheme outperforms separate or partial designs.The superiority of multi-view inference over single-view inference is also validated.
II. SYSTEM MODEL
The system combines multi-device radar sensing, local feature processing, wireless feature transmission, and centralized edge inference within a real-time task-oriented architecture. ISAC devices switch between sensing and communication modes while the server coordinates inference under a shared time budget.
- System workflow: Multiple ISAC devices sense the environment, locally process and quantize data into feature subsets, then transmit cascaded features to an edge server for inference.The server requests sensing and processing before receiving feature subsets over wireless links.
- System workflow: The edge server uses a trained machine-learning model to make real-time decisions from features collected by the ISAC devices.The permitted completion time for the real-time inference task is denoted as T.
- ISAC operation: The DFRC transceiver alternates flexibly between radar sensing and communication using a shared radio-frequency front-end in a time-division manner.Sensing uses FMCW up-ramp chirps, while communication uses digitally modulated constant-frequency carriers.
- Resource model: Each device has sensing, computation, and communication durations, with sensing and computation times treated as constant and communication bandwidth shared across devices.The communication time is T_c,k, and the total communication bandwidth is B.
- Network assumptions: The AP coordinates the devices and can acquire global channel state information while channels remain static over the short task duration.The device-server channel gain is denoted as H_c,k.
B. Radar Sensing and Feature Generation Model
The sensing pipeline converts radar echoes into feature elements by sampling, filtering, slow-time processing, and PCA extraction. The resulting features include target information alongside clutter and noise, with device feature subsets modeled as independent.
- Signal acquisition: Each sensing snapshot contains M chirps, and received radar signals are sampled into complex-valued data arranged across fast-time and slow-time dimensions.The fast-time dimension supports ranging, while slow-time processing estimates Doppler information.
- Signal processing: An SVD-based linear filter is applied to the received data matrix to mitigate clutter before feature extraction.The filter is parameterized by singular values and singular vectors of the sensing matrix.
- Feature extraction: Features are extracted in the slow-time dimension by vectorizing the processed data and applying PCA to obtain principal, uncorrelated feature elements.The principal eigenspace can be obtained during model training and broadcast to the devices.
- Feature model: Each feature element contains a desired ground-truth component contaminated by clutter and sensing noise, with clutter modeled as Gaussian when many independent reflection paths exist.The sensing feature is normalized by transmit radar sensing power.
- Feature model: Device k generates a feature subset containing N_k elements, and subsets from different devices are assumed independent because sensing areas do not overlap.This independence supports combining multi-device features for inference.
C. Quantization Model
The quantization model represents recovered features as distorted versions of locally generated features and uses discriminant gain to characterize classification separability. Higher quantization gain reduces recovered-feature distortion, while whole-vector gain aggregates independent feature contributions.
- Quantization: Each feature element is quantized locally, and the resulting feature subset determines the communication overhead required to transmit it to the server.The quantization distortion is approximated as additive Gaussian distortion.
- Quantization: The quantized feature uses a linear quantizer with quantization gain √Q_k and an approximate Gaussian distortion term d_k.The receiver recovers the quantized feature using the same quantizer model.
- Quantization: A larger quantization gain √Q_k produces lower quantization distortion in the recovered feature.The quantization model links bit allocation to the quality of the feature delivered to the server.
- Discriminant gain: Discriminant gain is adopted as the classification inference-accuracy metric and is derived from KL divergence between class-conditioned feature distributions.The metric is evaluated for arbitrary class pairs and averaged over all class pairs.
- Discriminant gain: The discriminant gain of the whole recovered feature vector is obtained by combining the gains of its independent feature elements across devices and subsets.The recovered feature vector is formed by cascading the device-level feature subsets.
III. PROBLEM FORMULATION & SIMPLIFICATION
The formulation maximizes total discriminant gain subject to latency, successful-transmission, and energy constraints. The resulting problem is non-convex because of its objective and selected constraints, while the paper frames the integration of sensing, computation, and communication as its systematic contribution.
- A. Problem Formulation: The objective maximizes total discriminant gain under latency, successful-transmission, and energy constraints.
- A. Problem Formulation: The formulation links the discriminant-gain objective to the three resource constraints while retaining the device-level sensing, communication, and energy variables.
- 1) Latency Constraint:: The latency constraint requires each device’s sensing, computation, and communication time to remain below the permitted real-time inference latency.Tr,k and Tm,k are constant sensing and computation times; Tc,k is allocated communication time.
- 2) Successful Transmission Constraint:: Successful transmission requires the mutual information between generated and recovered quantized feature subsets not to exceed the device channel capacity.The channel capacity depends on noise power, communication time, transmit power, and channel gain.
- 3) Energy Constraint:: Each ISAC device must keep its energy consumption below its energy threshold.The energy model includes sensing power, transmit power, constant sensing time, communication time, and constant computation energy.
- 3) Energy Constraint:: Problem (P1) is non-convex because both its objective function and Constraints (C2) and (C3) are non-convex.
- 3) Energy Constraint:: The work is presented as the first to enhance inference performance through systematic integration of sensing, computation, and communication.The cited comparison identifies prior work as studying discriminant-gain maximization through progressive feature transmission.
B. Problem Simplification
Variable transformations convert the original formulation into a problem with convex constraints and a non-concave sum-of-ratios objective. The sum-of-ratios method then introduces auxiliary variables and alternates convex optimization with auxiliary-variable updates.
- B. Problem Simplification: The transformed variables represent normalized sensing noise power, normalized quantization distortion, and communication energy consumption.
- B. Problem Simplification: Problem (P2) has convex constraints, but its summation-over-multiple-ratios objective is non-concave and therefore remains non-convex.
- B. Problem Simplification: The proposed ISCC scheme jointly allocates sensing and transmit power, time, and quantization bits.
- B. Problem Simplification: The sum-of-ratios procedure introduces auxiliary variables to derive a convex sum-of-weighted-distortion-minimization problem from (P2).The resulting convex problem is solved by alternating joint power-and-quantization allocation with communication-time allocation.
- B. Problem Simplification: Lemma 1 states that the objective of (P2) is a sum of quasi-linear ratios and can be optimally addressed using the sum-of-ratios method.
- B. Problem Simplification: Each weighted-distortion term scales the sum of sensing noise power and quantization distortion, making the resulting problem convex.
- B. Problem Simplification: Problem (P2) can be optimally solved by alternating between solving the convex subproblem under fixed auxiliary variables and updating those variables.
B. An Alternating Method for Solving (P3)
The alternating method solves the convex subproblem by deriving power, quantization, and energy allocations from KKT conditions. The resulting observations relate allocation levels to class structure, feature dimensionality, discriminant requirements, and channel-related parameters.
- B. An Alternating Method for Solving (P3): With fixed auxiliary variables, the convex subproblem admits closed-form solutions and low computational complexity.The method alternates between joint power-and-quantization allocation and communication-time allocation.
- B. An Alternating Method for Solving (P3): The Lagrangian formulation uses nonnegative multipliers for the energy and latency-related constraints before applying KKT conditions.
- B. An Alternating Method for Solving (P3): Larger class counts or required discriminant gains call for greater sensing power, while longer sensing time can reduce the required sensing power.
- B. An Alternating Method for Solving (P3): Larger class counts, feature-element counts, and required discriminant gains require greater quantization gain for finer-grained feature representations.
- B. An Alternating Method for Solving (P3): Larger centroid distances require smaller quantization gain because well-separated classes can be discriminated with lower-resolution representations.
- 1) Joint Power and Quantization Bits Allocation:: The optimal communication-power allocation is obtained from the communication-energy expression and solved using a primal-dual method.
- 1) Joint Power and Quantization Bits Allocation:: Algorithm 1 iteratively solves normalized sensing noise, quantization distortion, and communication energy, then updates the associated multipliers.
2) Communication Time Allocation:
Communication-time allocation is handled through a convex feasibility problem that finds the minimum required duration and updates device times according to latency feasibility and throughput gaps. The update rule yields smaller weighted distortion and enhanced discriminant gain.
- 2) Communication Time Allocation:: Given sensing noise, communication energy, and quantization distortion, the method determines communication times through a feasibility problem.
- 2) Communication Time Allocation:: The minimum required total duration T* is compared with permitted latency T to assess whether the current weighted-distortion setting is feasible.
- 2) Communication Time Allocation:: When T* > T, the current sensing, quantization, and communication-energy values are infeasible because the latency constraint cannot be satisfied.
- 2) Communication Time Allocation:: When T* < T, remaining time is reallocated across devices, with more time assigned to devices requiring more communication capacity.
- 2) Communication Time Allocation:: When T* = T, the current time allocation is optimal.
- 2) Communication Time Allocation:: Algorithm 2 combines primal-dual optimization for the feasibility problem with iterative communication-time updates.
- 2) Communication Time Allocation:: The communication-time update uses the throughput gap between each device’s required communication load and available channel capacity.
- 2) Communication Time Allocation:: Proposition 1 states that the time-update rule produces a smaller weighted distortion level and enhanced discriminant gain.
3) Alternating Algorithm for Solving (P3):
The alternating algorithm alternates between communication-time allocation and sensing, quantization, and energy variables to solve (P3). It monotonically decreases weighted distortion and optimally solves the convex subproblem.
- Convergence property: Alternating optimization produces monotonically decreasing weighted distortion for (P3).
- Convergence property: Because (P3) is convex, the alternating method can solve it optimally and has a linear convergence rate.
- Algorithm procedure: The algorithm initializes communication times and alternates between updating communication time and the remaining resource variables.The remaining variables include sensing noise power, quantization distortion, communication energy, and discriminant gains.
C. Solution to (P2)
The solution procedure applies the sum-of-ratios method together with an alternating algorithm to obtain the optimal ISCC allocation for (P2). Experiments evaluate the scheme using simulated human-motion sensing and two inference models.
- Solution method: The sum-of-ratios method and alternating algorithm jointly solve (P3), after which variable transformations yield the optimal solution of (P2).
- Solution method: Algorithm 4 solves (P2) using channel gains, auxiliary variables, and the alternating solution for sensing, quantization, energy, and communication-time variables.
- Simulation setup: Human-motion datasets contain four classes: child walking, child pacing, adult walking, and adult pacing.
- Simulation setup: SVM and MLP models are trained on 800 undistorted samples and tested on 200 distorted samples.
B. Inference Algorithms
The experiments compare power-aware, time-aware, quantization-aware, and optimal allocation schemes using SVM and MLP inference. Accuracy generally improves with discriminant gain, available resources, and participating ISAC devices, while model behavior differs by distortion level.
- Compared schemes: The optimal allocation jointly assigns all parameters, whereas baselines fix sensing power, communication time, or quantization bits before optimizing the remaining variables.
- Inference accuracy versus discriminant gain: Inference accuracy increases with discriminant gain for both the SVM and MLP models.
- Inference accuracy versus discriminant gain: When discriminant gain is large, the SVM outperforms the neural network because sensing and quantization distortion is small.
- SVM comparison: The optimal allocation scheme outperforms the other three schemes for SVM inference as energy thresholds and permitted latency increase.
- MLP comparison: The optimal allocation scheme also achieves the best MLP performance, while additional permitted latency eventually stops improving the power-aware scheme.
- Number of ISAC devices: Increasing the number of ISAC devices improves inference accuracy for both models because more features enlarge the feature space and discriminant gains.
- Model comparison: The SVM outperforms the MLP in the reported setting because neural-network training is overfitting.
- Overall findings: The experiments report that the optimal ISCC scheme has the best performance and verify the theoretical analysis.
APPENDIX
The appendix establishes that (P2) has a convex feasible region and a sum-of-ratios objective structure, supporting optimal solution by the sum-of-ratios method. It also explains the alternating structure of (P3).
- Convexity of (P2): The objective of (P2) can be written as a sum of quasi-linear ratios with constant coefficients.
- Convexity of (P2): Because the feasible region is convex, (P2) can be optimally solved using the sum-of-ratios method.
- Convexity of (P2): The latency constraint is linear, while the remaining constraints form convex sets through concavity, positive second derivatives, and linear transformations.
- Alternating solution of (P3): For (P3), increasing communication time reduces the communication energy required by each device.
- Alternating solution of (P3): Lower communication energy can reduce sensing noise, while the objective increases with sensing noise through the defined function and enhances discriminant gain.