Imprecise Deep Networks for Uncertain Image Classification.

Chuanqi Liu,Zuowei Zhang,Zechao Liu,Liangbo Nine,Zhunga Liu
DOI: https://doi.org/10.1007/978-3-031-67977-3_3
2024-01-01
Abstract:Deep learning techniques have been successfully applied in image classification tasks. Still, data uncertainty is hindering the demand for higher performance. Existing techniques can only suppress the uncertainty caused by one reason, considered a passive strategy. Here, we introduce an open imprecise deep network (ImpNN) framework for image classification to actively handle data uncertainty. The ImpNN can model and reason data uncertainty-caused imprecision by using meta-class, defined as the union of different specific categories to constrain and improve the network performance. In addition, ImpNN characterizes and exploits imprecision by introducing two new loss functions (Imprecision loss function and Denoising loss function), which separately contribute to exploit the mined imprecision and alleviate the side effect of imprecision. We employ several typical networks to theoretically and experimentally analyze the ImpNN, which presents better performance compared to other methods based on open datasets. Experimental evaluations also show that our ImpNN can characterize imprecise information in results, potentially for cautious decision-making applications.
What problem does this paper attempt to address?