Paper Title
Joint Policy Learning for Adaptive Early-Exit Inference in Resource-Constrained Deep Learning

Abstract
Deep neural networks achieve high accuracy in image classification but require significant computational resources and energy, posing challenges for resource-constrained devices like mobile phones and embedded systems. Early-exit architectures mitigate this by allowing easy inputs to exit from intermediate layers, reducing average inference cost. We propose a novel approach to learn an early-exit policy using lightweight policy networks trained jointly with the classifier. Our method augments a MobileNetV2 backbone with multiple exit branches and employs a learned policy to dynamically decide whether to output a prediction early or continue to deeper layers. A joint loss function balances classification accuracy and energy efficiency by penalizing unnecessary computation. Experiments on CIFAR-10, STL-10, and TinyImageNet-200 show that our adaptive model maintains near-baseline accuracy while reducing computation (FLOPs) and estimated energy by 30–35%. We provide detailed results, including accuracy/FLOPs/energy trade-offs, exit distribution analysis, and ablation studies on policy parameters. This framework offers a promising solution for energy-efficient deep learning in resource-constrained environments. To facilitate reproducibility, we will release the code upon acceptance. Keywords - Joint Policy Learning, Early-Exit Networks, Adaptive Inference, Efficient Inference, Edge AI.