Optimized Cooperative Inference for Energy-Efficient and Low-Latency Mobile Edge Computing

To overcome the limitations of standalone inference on edge devices or servers, we propose a cooperative inference method for mobile edge computing (MEC) systems. Using dual confidence thresholds on a small neural network (NN) at the edge, ambiguous images are filtered and sent to a larger NN on the server for reevaluation. We evaluate the method's accuracy, delay, and energy consumption, accounting for confidence score distributions that could trigger false alarms. A joint optimization problem is formulated to minimize delay and energy consumption by selecting optimal confidence thresholds, transmit power, and duty cycle while ensuring accuracy. Experimental results show that this approach significantly reduces delay and energy consumption while achieving higher accuracy than device-only inference and lower costs than server-only inference in various MEC scenarios.