Cross-Layer Neural Network Optimization for Balancing Accuracy, Speed, and Energy Consumption in Real-World Applications

Abstract

As machine learning applications proliferate across mobile, embedded, and edge devices, there is a pressing need to optimize neural networks not only for accuracy but also for computational speed and energy efficiency. Traditional approaches that focus on single-layer optimizations often fall short in meeting the constraints of real-world applications. This paper presents a cross-layer optimization framework that integrates algorithmic, architectural, and hardware-level adaptations to holistically balance accuracy, latency, and energy consumption. The proposed framework enables neural networks to dynamically reconfigure their behavior based on runtime constraints, leveraging techniques such as layer fusion, quantization-aware training, memory hierarchy reorganization, and adaptive activation pruning. Results from empirical evaluations on diverse benchmarks demonstrate that our method achieves up to 40% energy reduction and 30% latency improvement with negligible accuracy degradation. These findings pave the way for more sustainable and deployable AI systems in edge computing and mobile inference.