Call for Papers: *IEEE Transactions on Neural Networks and Learning Systems*
*Special Issue on **Deep Representation and Transfer Learning for Smart and Connected Health* *Important Dates* 31 March 2019 – Deadline for manuscript submission 30 June 2019 – Reviewer’s comments to authors 31 August 2019 – Submission of revised papers 31 October 2019 – Final decision of acceptance 30 November 2019 – Camera-ready papers December 2019-February 2020 – Tentative publication date *Aims and Scope**:* Deep neural networks have proven to be efficient learning systems for supervised and unsupervised tasks in a wide range of challenging applications. However, learning complex data representations using deep neural networks can be difficult due to problems such as lack of data, exploding or vanishing gradients, high computational cost, or incorrect parameter initialization, among others. Transfer Learning (TL) can facilitate the learning of data representations by taking advantage of transferable features learned by a model in a source domain, and adapting the model to a new domain. This approach has demonstrated to produce better generalization performance than random weight initialization, and has produced state-of-the-art results in signal and visual processing tasks. Accordingly, emerging and challenging domains, such as smart and connected health (SCH), can benefit from new theoretical advancements in representation and transfer learning (RTL) methods. One of the main advantages of TL is its potential to be applied in a wide range of domains and for different learning tasks. For instance, in facial affect recognition, the representations learned by a deep model trained to recognize faces in an unsupervised fashion can be employed and improved by a second model to perform emotion recognition in supervised manner. Nonetheless, learning data representations that provide a good degree of generalization performance remains a challenge. This is due to issues such as the inherent trade-off between retaining too much information from the input and learning universal features. Similarly, despite the obvious advantages of TL, effective use of parameters learned by a given model in a different domain is a challenge, particularly when there is limited data in the target domain. This challenge increases when the joint distribution of the input features and output labels is different in the target domain. In addition, determining how to reject unrelated information or remove dataset bias during TL is yet to be solved. Other limitations are caused by lack of existing theoretical approaches in RTL capable of explaining or interpreting the learning process of deep models, or determining how to best learn a set of data representations that are ideal for a given task, whether in a regression or classification problem. Therefore, new n theoretical mechanisms and algorithms are required to improve the performance and learning process of deep neural networks. Despite these constraints, RTL will play an essential role in building the next generation of intelligent systems designed to assists humans with their daily needs. Consequently, domains of great interest to human society, such as SCH, will benefit from new advancements in RTL. For instance, one of the main challenges in designing effective SCH applications is overcoming the lack of labelled data. RTL can overcome this limitation by training a model to learn universal data representations on larger corpora in a different domain, and then adapting the model for use in a SCH context. Similarly, RTL can be used in conjunction with generative adversarial networks to overcome class imbalance problems by generating new healthcare-related data, which can also be used to improve the generalization performance of deep models in SCH applications. Furthermore, RTL can be used to initialize and improve the learning of deep reinforcement learning models designed for continuous learning in patient-centered cognitive support systems, among others. Nonetheless, the use of RTL in designing SCH applications requires overcoming problems such as dataset bias or neural network co-adaptation. This special issue on Deep Representation and Transfer Learning for Smart and Connected Health invites researchers and practitioners to present novel contributions addressing theoretical aspects of representation and transfer learning. The special issue will provide a collection of high quality research articles addressing theoretical work aimed to improve the generalization performance of deep models, as well as new theory attempting to explain and interpret both concepts. State-of-the-art works on applying representation and transfer learning for developing smart and connected health intelligent systems are also very welcomed. Topics of interest for this special issue include but are not limited to: *Theoretical Methods:* · Distributed representation learning; · Transfer learning; · Invariant feature learning; · Domain adaptation; · Neural network interpretability theory; · Deep neural networks; · Deep reinforcement learning; · Imitation learning; · Continuous domain adaptation learning; · Optimization and learning algorithms for DNNs; · Zero and one-shot learning; · Domain invariant learning; · RTL in generative and adversarial learning; · Multi-task learning and Ensemble learning; · New learning criteria and evaluation metrics in RTL; * Application Areas: * · Health monitoring; · Health diagnosis and interpretation; · Early health detection and prediction; · Virtual patient monitoring; · RTL in medicine; · Biomedical information processing; · Affect recognition and mining; · Health and affective data synthesis; · RTL for virtual reality in healthcare; · Physiological information processing; · Affective human-machine interaction; *Guest Editors* *Vasile Palade*, Coventry University, UK *Stefan Wermter*, University of Hamburg, Germany *Ariel Ruiz-Garcia*, Coventry University, UK *Antonio de Padua Braga*, University of Minas Gerais, Brazil *Clive Cheong Took*, Royal Holloway (University of London), UK *Submission Instructions * 1. Read the Information for Authors at https://cis.ieee.org/publications/t-neural-networks-and-learning-systems. 2. Submit your manuscript at the TNNLS webpage ( http://mc.manuscriptcentral.com/tnnls) and follow the submission procedure. Please, clearly indicate on the first page of the manuscript and in the cover letter that the manuscript is submitted to this special issue. Send an email to the guest editors Vasile Palade ( vasile.pal...@coventry.ac.uk) and Ariel Ruiz-Garcia ( ariel.ruiz-gar...@coventry.ac.uk) with subject “TNNLS special issue submission” to notify about your submission. 3. Early submissions are welcome. We will start the review process as soon as we receive your contributions. For any other questions please contact Ariel Ruiz-Garcia ( ariel.ruiz-gar...@coventry.ac.uk).
_______________________________________________ uai mailing list uai@ENGR.ORST.EDU https://secure.engr.oregonstate.edu/mailman/listinfo/uai
