Lowering mutual coherence between receptive fields in convolutional neural networks. Issue 6 (1st March 2019)

• Record Type: Journal Article
• Authors: Amini, S.; Ghaemmaghami, S.
• Is Part Of: Electronics letters. Volume 55:Issue 6(2019)
• Page Start: 325
• Page End: 327
• Publication Date: 2019-03-01
• Subjects: signal reconstruction -- sparse matrices -- compressed sensing -- convolutional neural nets; mutual coherence -- receptive fields -- convolutional neural networks -- low‐coherence dictionaries -- sparse signal processing -- low‐coherence attribute -- low‐coherence convolutional neural network -- signal recovery; Electronics -- Periodicals; 621.381
• DOI: 10.1049/el.2018.7671
• Languages: English
• ISSNs: 0013-5194

Abstract

Abstract : It has been shown that more accurate signal recovery can be achieved with low‐coherence dictionaries in sparse signal processing. In this Letter, the authors extend the low‐coherence attribute to receptive fields in convolutional neural networks. A new constrained formulation to train low‐coherence convolutional neural network is presented and an efficient algorithm is proposed to train the network. The resulting formulation produces a direct link between the receptive fields of a layer through training procedure that can be used to extract more informative representations from the subsequent layers. Simulation results over three benchmark datasets confirm superiority of the proposed low‐coherence convolutional neural network over the unconstrained version.