This repo provides official code for the paper "Enhancing Adaptive Deep Networks for Image Classification via Uncertainty-aware Decision Fusion" publised in ACM Multimedia 2024. The full version paper is available at the link.
When computational resources are limited, deep networks cannot be used. Currently, several adaptive network architectures with multi-head classifiers are proposed to alleviate the issue of image classification under limited computational resources. However, the knowledge of classifiers in different depths are not fully utilized. This work investigates how to improve the accuracy of different classifiers without significantly increasing computational resources.
Key words: limited computing resources, image classification, uncertainty-aware fusion, ensemble learning, adaptve deep networks, multi-head classifiers
(1) Create a "data" folder and place the downloaded CIFAR10, CAIFAR100, ImageNet100 and ImageNet1000 datasets into it.
(2) Customize the parameters in "main_CIFAR_train.py" or "main_ImageNet_train.py" and run to train the RANet based on the GCDM framework.
(3) Run "main_CIFAR_test.py" or "main_ImageNet_test.py" and apply our proposed CDM module to reproduce the results in the paper across various datasets.
If you find this repo helpful, please cite our paper.
@inproceedings{zhang2024enhancing,
title={Enhancing Adaptive Deep Networks for Image Classification via Uncertainty-aware Decision Fusion},
author={Zhang, Xu and Xie, Zhipeng and Yu, Haiyang and Wang, Qitong and Wang, Peng and Wang, Wei},
booktitle={ACM Multimedia 2024}}