Surrogate network-based sparseness hyper-parameter optimization for deep expression recognition. (March 2021)

• Record Type: Journal Article
• Authors: Xie, Weicheng; Chen, Wenting; Shen, Linlin; Duan, Jinming; Yang, Meng
• Is Part Of: Pattern recognition. Volume 111(2021)
• Publication Date: 2021-03
• Subjects: Expression recognition -- Deep sparseness strategies -- Hyper-parameter optimization -- Surrogate network -- Heuristic optimizer; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107701
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A new iterative framework for the hyper-parameter optimization in deep sparseness strategies is proposed to adapt hyper-parameter setting to different databases in FER. A simplified network is deployed to surrogate the original network for hyper-parameter optimization, where Euclidean losses with unilateral back propagation are introduced to approximate the original network. The proposed algorithm can automatically adapt deep network metrics to different databases with a reasonable time complexity. The hyper-parameter optimization algorithm achieved competitive performances on six public benchmark expression databases. Abstract: For facial expression recognition, the sparseness constraints of the features or weights can improve the generalization ability of a deep network. However, the optimization of the hyper-parameters in fusing different sparseness strategies demands much computation, when the traditional gradient-based algorithms are used. In this work, an iterative framework with surrogate network is proposed for the optimization of hyper-parameters in fusing different sparseness strategies. In each iteration, a network with significantly smaller model complexity is fitted to the original large network based on four Euclidean losses, where the hyper-parameters are optimized with heuristic optimizers. Since the surrogate network uses the same deep metrics and embeds the same hyper-parameters as the original network, the optimized hyper-parameters are then used for the training of the original deep network in the next iteration. While the performance of the proposed algorithm is justified with a tiny model, i.e. LeNet on the FER2013 database, our approach achieved competitive performances on six publicly available expression datasets, i.e., FER2013, CK+, Oulu-CASIA, MMI, AFEW and AffectNet.