Neural network optimisation: quantisation techniques, practical benchmarking, and deployment recommendations

Abstract

The paper addresses the deployment of deep convolutional neural networks under the memory and energy limitations typical of mobile hardware, focusing on quantisation as a practical optimisation technique. The author performs a systematic empirical comparison of five convolutional architectures (MobileNetV2, ResNet-50, DenseNet-121, EfficientNet-B3, EfficientNet-B4), each converted with TensorFlow Lite Converter into three numerical-precision formats — FP32 baseline, FP16 and INT8 — and benchmarked on a physical iPhone 13 device (Apple A15 Bionic) through a purpose-built React Native 0.82 application using the react-native-fast-tflite library. Classification accuracy is assessed on a 500-image ImageNet-v2 subset. The results demonstrate that FP16 conversion halves model size while keeping Top-1 accuracy loss below 0.3% for every architecture, whereas INT8 reduces file size by 71–74% and improves inference latency by up to 2.8 times for mobile-friendly networks, but yields no consistent latency gain for EfficientNet-B3/B4 because part of their computation graph falls back to CPU execution within the TensorFlow Lite Metal delegate. Three factors governing quantisation effectiveness are identified: the structural compatibility of an architecture with the hardware delegate, the representativeness of the calibration dataset, and the delegate's operator coverage. Based on these findings, the author formulates differentiated, evidence-based deployment recommendations for mobile AI developers, organised by architecture family and use-case requirements, and proposes a reusable conversion-and-benchmarking methodology applicable beyond the five architectures studied.