Authors - Moncef Zarrouk, Abdelmajid Bessate, Faissal El Bouanani Abstract - Edge artificial intelligence systems require inference mechanisms that jointly account for predictive accuracy and resource consumption. Compact edge models provide low-cost and low-latency inference, but they may be less reliable on difficult inputs, whereas larger cloud models are more accurate but expensive to invoke for every sample. In this paper, we propose a constrained optimization framework for calibration-aware adaptive edge–cloud inference. The edge model first produces a local prediction and a confidence score. This confidence is post-calibrated by scalar temperature scaling, and a threshold gate then either accepts the local prediction or defers the input to a cloud model. The threshold is selected by solving an empirical constrained optimization problem that maximizes accuracy under an operator-specified of- load budget. We show that temperature scaling preserves the edge-model decision, that the empirical offload rate is monotone in the threshold, and that the finite empirical threshold-selection problem can be solved exactly by enumerating the gate partitions induced by the observed confidence scores, correctly handling the strict deferral boundary. Experiments on Fashion-MNIST show that constrained deferral recovers most of the edge–cloud accuracy gap while offloading only a controlled fraction of inputs. Moreover, an additive-noise stress test shows that calibration substantially reduces confidence miscalibration when the edge model be- comes unreliable, confirming that calibration mainly contributes reliabil- ity, while selective deferral drives the accuracy gain.
