Convolutional neural networks analysed via inverse problem theory and sparse representations

Download

index.pdf

Date

2019-04-01

Author

Tarhan, Cem
Akar, Gözde

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

212
views

0
downloads

Inverse problems in imaging such as denoising, deblurring, superresolution have been addressed for many decades. In recent years, convolutional neural networks (CNNs) have been widely used for many inverse problem areas. Although their indisputable success, CNNs are not mathematically validated as to how and what they learn. In this study, the authors prove that during training, CNN elements solve for inverse problems which are optimum solutions stored as CNN neuron filters. They discuss the necessity of mutual coherence between CNN layer elements in order for a network to converge to the optimum solution. They prove that required mutual coherence can be provided by the usage of residual learning and skip connections. They have set rules over training sets and depth of networks for better convergence, i.e. performance. They have experimentally validated theoretical assertions.

Subject Keywords

Signal Processing, Electrical and Electronic Engineering

URI

https://hdl.handle.net/11511/37389

Journal

IET SIGNAL PROCESSING

DOI

https://doi.org/10.1049/iet-spr.2018.5220

Collections

Department of Electrical and Electronics Engineering, Article

Suggestions

OpenMETU
Core

Properly Handling Complex Differentiation in Optimization and Approximation Problems Candan, Çağatay (Institute of Electrical and Electronics Engineers (IEEE), 2019-03-01) Functions of complex variables arise frequently in the formulation of signal processing problems. The basic calculus rules on differentiation and integration for functions of complex variables resemble, but are not identical to, the rules of their real variable counterparts. On the contrary, the standard calculus rules on differentiation, integration, series expansion, and so on are the special cases of the complex analysis with the restriction of the complex variable to the real line. The goal of this lect...
Exact Relation Between Continuous and Discrete Linear Canonical Transforms Öktem, Sevinç Figen (Institute of Electrical and Electronics Engineers (IEEE), 2009-08-01) Linear canonical transforms (LCTs) are a family of integral transforms with wide application in optical, acoustical, electromagnetic, and other wave propagation problems. The Fourier and fractional Fourier transforms are special cases of LCTs. We present the exact relation between continuous and discrete LCTs (which generalizes the corresponding relation for Fourier transforms), and also express it in terms of a new definition of the discrete LCT (DLCT), which is independent of the sampling interval. This p...
Joint spatial and temporal channel-shortening techniques for frequency selective fading MIMO channels Toker, Canan; Chambers, JA; Baykal, Buyurman (2005-02-01) It is well understood that the maximum likelihood estimator is a powerful equalisation technique for frequency selective fading channels, and in particular for MIMO systems. The complexity of this estimator, however, grows exponentially with the number of users and multipath taps, hence limiting the use of this algorithm in MIMO systems. In the paper, the authors propose a joint spatial and temporal channel-shortening filter as a pre-processor to reduce significantly the complexity of a maximum likelihood e...
Transmit beamformer design with a PAPR constraint to trade-off between beampattern shape and power efficiency Çayır, Ömer; Candan, Çağatay (Elsevier BV, 2020-04-01) This study examines the effect of peak-to-average power ratio (PAPR) constraint on the transmit beamformer design problem with the goal of establishing a trade-off between the power efficiency (maximizing the average transmitted power) and other metrics such as the power level fluctuation in mainlobe, peak-sidelobe level (PSL), etc. Typically, unimodular weights are utilized in transmit beamforming to maximize the average transmitted power. Yet, unimodular weights maximize the power efficiency at the expens...
Semi-supervised dimension reduction approaches integrating global and local pattern information Sakarya, Ufuk (Springer Science and Business Media LLC, 2019-02-01) Dimension reduction is an important research area in pattern recognition. Use of both supervised and unsupervised data can be an advantage in the case of lack of labeled training data. Moreover, use of both global and local pattern information can contribute classification performances. Therefore, four important primary components are essential to design a well-performed semi-supervised dimension reduction approach: global pattern modeling by a supervised manner, local pattern modeling by a supervised manne...

Citation Formats

C. Tarhan and G. Akar, “Convolutional neural networks analysed via inverse problem theory and sparse representations,” IET SIGNAL PROCESSING, pp. 215–223, 2019, Accessed: 00, 2020. [Online]. Available: https://hdl.handle.net/11511/37389.