Adaptive Depth Neural Networks for Energy-Efficient Edge Intelligence

I. ABSTRACT Edge devices such as IoT sensors, mobile processors, and embedded systems operate under strict constraints in computation, memory, and power consumption, making traditional deep learning models unsuitable for real-time deployment. Conventional fixed-depth neural networks process every input through all layers, leading to redundant computations, increased latency, and higher energy usage— particularly for simple inputs that do not require full processing depth. To overcome these challenges, this project introduces Smart Depth, an energy-efficient dynamic-depth neural network architecture optimized for edge intelligence. The system employs lightweight gating mechanisms and confidence based early-exit classifiers to adaptively determine the necessary inference depth based on input complexity and prediction certainty. This conditional and adaptive computation strategy significantly reduces power consumption, lowers inference latency, and preserves model accuracy. Experimental results confirm that Smart Depth delivers substantial improvements in energy efficiency, responsiveness, and overall performance, making it highly suitable for IoT devices, mobile AI platforms, and embedded real-time applications.