As a result of radical advances at the hardware and algorithmic level and due to the increase in the availability of data, the last decade has been marked by the tremendous success of deep learning in various tasks in signal analysis and, more recently, reconstruction and processing. Currently, deep-neural-network models constitute the state of the art in analysis, reconstruction and generative tasks in different applications involving various types of data, including image/video, text, and audio data, networked data (IoT data, social media data), and biomedical and bioinformatics data. A fundamental limiting factor for deep learning is that its theoretical understanding remains underdeveloped, which in turn translates into a lack of principled methods to design such architectures. In response to these challenges, this special issue invites contributions revolving around: the theoretical foundations of deep learning, with particular interest in principles linked or derived from signal processing; the design of interpretable deep neural networks through such foundations; and the application of interpretable deep learning in various tasks in signal processing, reconstruction, generation and analysis.
The first theme refers to contributions on theoretical frameworks and foundations around:
- the dynamics of optimization algorithms used to train neural networks
- the impact of the architectural choices, such as the number and type of layers, the number of neurons per layer, and the effect of regularization strategies
- the exploitation of structure in the data, such as sparsity, low-rank and side information
The second theme refers to contributions on principled recipes to design deep neural networks, particularly but not limited to:
- interpretable feedforward, convolutional, and recurrent architectures
- geometric deep learning architectures, including graph convolutional, and deep graphical models
- explainable generative models (e.g., generative adversarial networks, variational autoencoders)
- interpretable multimodal deep learning architectures that merge diverse but correlated data
The third theme refers to the use of interpretable deep learning to solve problems in data
- data classification: recognition, semantic segmentation, etc.
- data reconstruction: superresolution, denoising, deconvolution, compressed sensing, matrix completion, and inverse problems in general
- data analysis, dimensionality reduction and clustering
The issue is open to applications in different domains, from traditional disciplines – such as remote sensing, medical imaging, human-machine interaction, and computational imaging – to emerging fields – like autonomous vehicles, social media, Internet-of-Things, precision agriculture and personalized medicine.
Submission is permitted only if the paper has not been submitted, accepted, published, or copyrighted in another journal. Papers that have been published in conference and workshop proceedings may be submitted for consideration provided that (i) the authors cite their earlier work; (ii) the papers are not identical; and (iii) the journal publication includes novel elements (e.g., more comprehensive experiments). For submission information, please refer to the submission guidelines at https://asp-eurasipjournals.springeropen.com/submission-guidelines.
Initial Paper Submission: September 30, 2019
1st Review Completed: December 15, 2019
Revised Manuscript Due: February 15, 2020
2nd Review Completed: April 15, 2020
Final Manuscript Due: May 15, 2020
Lead Guest Editor:
Nikos Deligiannis, Vrije Universiteit Brussel, Belgium
Saikat Chatterjee, KTH, Sweden
Monika Dörfler, University of Vienna, Austria
Raja Giryes, Tel Aviv University, Israel
Zhanyu Ma, Beijing University of Posts and Telecommunications, China