Abstract
Least mean square error reconstruction for the self-organizing network (Lmser) was proposed in 1991, featured by a bidirectional architecture with several built-in natures. In this paper, we developed Lmser into CNN based Lmser (CLmser), highlighted by new findings on strengths of two major built-in natures of Lmser, namely duality in connection weights (DCW) and duality in paired neurons (DPN). Shown by experimental results on several real benchmark datasets, DCW and DPN bring to us relative strengths in different aspects. While DCW and DPN can both improve the generalization ability of the reconstruction model on small-scale datasets and ease the gradient vanishing problem, DPN plays the main role. Meanwhile, DPN can form shortcut links from the encoder to the decoder to pass detailed information, so it can enhance the performance of image reconstruction and enables CLmser to outperform some recent state-of-the-art methods in image inpainting with irregularly and densely distributed point-shaped masks.
Similar content being viewed by others
References
Ballard DH (1987) Modular learning in neural networks. In: Proceedings of the sixth national conference on artificial intelligence—volume 1, AAAI’87, pp 279–284
Bertalmio M, Sapiro G, Caselles V, Ballester C (2000) Image inpainting. In: Proceedings of the 27th annual conference on computer graphics and interactive techniques, SIGGRAPH ’00. ACM Press/Addison-Wesley Publishing Co., New York, NY, USA, pp 417–424
Bottou L (2010) Large-scale machine learning with stochastic gradient descent. In: Proceedings of COMPSTAT’2010. Springer, pp 177–186
Bourlard H, Kamp Y (1988) Auto-association by multilayer perceptrons and singular value decomposition. Biol Cybern 59(4–5):291–294
Coates A, Ng A, Lee H (2011) An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics, pp 215–223
Han K, Wen H, Zhang Y, Fu D, Culurciello E, Liu Z (2018) Deep predictive coding network with local recurrent processing for object recognition. In: Advances in neural information processing systems, pp 9201–9213
He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770–778
Heusel M, Ramsauer H, Unterthiner T, Nessler B, Hochreiter S (2017) Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in neural information processing systems, pp 6626–6637
Hinton GE, Salakhutdinov RR (2006) Reducing the dimensionality of data with neural networks. Science 313(5786):504–507
Huang G, Liu Z, Weinberger KQ (2016) Densely connected convolutional networks. CoRR arXiv:1608.06993
Isola P, Zhu JY, Zhou T, Efros AA (2017) Image-to-image translation with conditional adversarial networks. In: CVPR
Karras T, Aila T, Laine S, Lehtinen J (2018) Progressive growing of GANs for improved quality, stability, and variation. In: International conference on learning representations. https://openreview.net/forum?id=Hk99zCeAb
Kingma DP, Ba J (2014) Adam: a method for stochastic optimization. CoRR arXiv:1412.6980
Krizhevsky A, Hinton G et al (2009) Learning multiple layers of features from tiny images
LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document recognition. Proc IEEE 86(11):2278–2324
Liu G, Reda FA, Shih KJ, Wang T, Tao A, Catanzaro B (2018) Image inpainting for irregular holes using partial convolutions. CoRR arXiv:1804.07723
Liu Z, Luo P, Wang X, Tang X (2015) Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision, pp. 3730–3738
Mao X, Shen C, Yang YB (2016) Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections. In: Advances in neural information processing systems, pp 2802–2810
Ronneberger O, Fischer P, Brox T (2015) U-net: convolutional networks for biomedical image segmentation. In: International conference on medical image computing and computer-assisted intervention. Springer, pp 234–241
Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-propagating errors. Nature 323(6088):533–536
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP et al (2004) Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 13(4):600–612
Xiao H, Rasul K, Vollgraf R (2017) Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747
Xu L (1991) Least MSE reconstruction for self-organization: (i)&(ii). In: Proceedings of 1991 international joint conference on neural networks, pp 2363–2373
Xu L (1993) Least mean square error reconstruction principle for self-organizing neural-nets. Neural Netw 6(5):627–648
Xu L (2019) An overview and perspectives on bidirectional intelligence: Lmser duality, double IA harmony, and causal computation. IEEE/CAA J Autom Sin 6(4):865–893
Zeng Y, Fu J, Chao H, Guo B (2019) Learning pyramid-context encoder network for high-quality image inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1486–1494
Zheng C, Cham TJ, Cai J (2019) Pluralistic image completion. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1438–1447
Zhou B, Lapedriza A, Khosla A, Oliva A, Torralba A (2017) Places: a 10 million image database for scene recognition. IEEE Trans Pattern Anal Mach Intell 40(6):1452–1464
Acknowledgements
This work was supported by National Science and Technology Innovation 2030 Major Project (2018AAA0100700) of the Ministry of Science and Technology of China, and SJTU Medical Engineering Cross-cutting Research Foundation (ZH2018ZDA07), as well as ZhiYuan Chair Professorship Start-up Grant (WF220103010) from Shanghai Jiao Tong University.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
A Implementations
For PConv [16], we use an implementation in PyTorch.Footnote 1 For PEN-Net [26], we use their official implementationFootnote 2 and follow all their original settings. For PICNet [27], we use their official implementationFootnote 3 and follow all the settings in their original work except that we start training from the provided pre-trained models and training for 1,000,000 iterations for CelebA-HQ and 2,000,000 iterations for Places2.
B Additional Results on Reconsturction Capability
To further investigate the reconstruction capability of CLmser-n and AE, we conduct experiments on Places2 dataset. We train the two models on 180,346 samples randomly chosen from the train set of Places2 and test them on 10,000 images which are randomly chosen from the test set of Places2. The results are listed in Table 8.
The results show that the design of the symmetric weights (DCW) does not affect the reconstruction performance.
C Qualitative Examples
See Fig. 8.
Rights and permissions
About this article
Cite this article
Huang, W., Tu, S. & Xu, L. Deep CNN Based Lmser and Strengths of Two Built-In Dualities. Neural Process Lett 54, 3565–3581 (2022). https://doi.org/10.1007/s11063-020-10341-5
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11063-020-10341-5