Description: Deep learning models and techniques are nowadays the leading approaches to face complex machine learning and pattern recognition problems, especially when considering perceptual tasks such as speech and image understanding. The adoption of deep architectures, comprising multiple, adaptable, processing layers, has recently allowed significant improvements in performance for these type of tasks. Both unsupervised and supervised approaches for training deep architectures have been empirically explored, also thanks to the adoption of parallel computer facilities such as GPUs or CPU clusters. Despite of that, there is a limited understanding of why deep architectures work so well and on how to design computationally efficient and effective training algorithms. A natural source of inspiration for a better understanding of these issues is the study of human brain, where deep structures are now well recognized and pervasive (e.g. human visual recognition requires the activation of a hierarchy of processing stages and pathways.) This special session focuses on all aspects of deep architectures and deep learning, with a particular emphasis on understanding fundamental principles. Because of that, it aims to bring together leading scientists in deep learning and related areas within neuroscience, machine learning, artificial intelligence, mathematics, and statistics. Topics of interest to the special session include, but are not limited to: Theoretical results on representation and learning in natural or artificial deep architectures; Theoretical and/or empirical analysis of specific natural or artificial deep architectures or algorithms; Innovative deep architectures/algorithms for data representation and analysis, including both supervised methods like deep convolution networks and unsupervised ones like stacked auto-encoders and deep Boltzmann machines; Design and/or analysis of recurrent and recursive architectures for processing of sequences and more general data structures; Applications of deep learning in data representation and analysis, including recognition, understanding, detection, segmentation, retrieval, restoration, super-resolution, and compression; Deep learning algorithms that efficiently handle large-scale data; Deep learning software and hardware architectures for applications. Special Session on Deep Learning @ ESANN 2016

Description: Deep learning has demonstrated its capability for many vision problems, such as face detection and recognition, image classification, etc. It is expected that this technique can benefit the area of medical image analysis, as well as imaging-based translational medicine. Though a few pioneering works can be found in the literature, there are still a lot of unresolved issues when applying deep learning for medical images. The goal of special session is to present works that focus on the design and use of deep learning in medical image analysis as well as imaging-based translational medical studies. This special session is going to set the trends and identify the challenges of the use of deep learning methods in the field of medical image. Meanwhile, it is expected to increase the connection between software developers, specialist researchers and applied end-users from diverse fields. Topics of interest to the special session include, but are not limited to: Image descriptor and feature extraction; Image super-resolution; Image reconstruction; Image registration; Image segmentation and labeling; Computer-assisted lesion detection; Computer-assisted diagnosis; Deep learning model selection; Meta-heuristic techniques for fine-tuning parameter in deep learning-based architectures; Other related translational medical applications. Organized by Qian Wang, Jun Shi, Shihui Ying, Manhua Liu and Yonghong Shi Special Session Web Site

Description: Deep learning models and techniques are becoming more and more the computational tool of choice when facing difficult applicative problems, such as speech and image understanding. The reason for this huge interest in deep learning is due to the fact that their adoption leads to human (and, in some cases, super-human) performances. These successes, however, have been mainly obtained on empirical basis, often thanks to the computational power provided by parallel computer facilities such as GPUs or CPU clusters. Although some recent works have addressed deep learning from a theoretical perspective, still there is a limited understanding of why deep architectures work so well and on how to design computationally efficient and effective training This special session aims to gather together leading scientists in deep learning and related areas within computational intelligence, neuroscience, machine learning, artificial intelligence, mathematics, and statistics, interested in all aspects of deep architectures and deep learning, with a particular emphasis on understanding fundamental principles. Topics of interest to the special session include, but are not limited to: Theoretical results on representation and learning in natural or artificial deep architectures; Theoretical and/or empirical analysis of specific natural or artificial deep architectures or algorithms; Innovative deep architectures/algorithms for data representation and analysis, including both supervised methods like deep convolution networks and unsupervised ones like stacked auto-encoders and deep Boltzmann machines; Design and/or analysis of recurrent and recursive architectures for processing of sequences and more general data structures; Applications of deep learning in data representation and analysis, including recognition, understanding, detection, segmentation, retrieval, restoration, super-resolution, and compression; Deep learning algorithms that efficiently handle large-scale data. Alessandro Sperduti, Univ. Padova (Italy), Jose C. Principe, University of Florida (USA), Plamen Angelov, Lancaster University (UK). Special Session Web Site

Description: Conventional multimedia understanding is usually built on top of handcrafted features, which are often much restrictive in capturing complex multimedia content. Recent progress on deep learning opens an exciting new era, placing multimedia understanding on a more rigorous foundation with automatically learned representations to model the multimodal data and the cross-media interactions. Existing studies have revealed promising results that have greatly advanced the state-of-the-art performance in a series of multimedia research areas, from the multimedia content analysis, to modeling the interactions between multimodal data, to multimedia content recommendation systems, to name a few here. This special session aims to provide a forum for the presentation of recent advancements in deep learning research that directly concerns the multimedia community. For multimedia research, it is especially important to develop deep learning methods to capture the dependencies between different genres of data, building joint deep representation for diverse modalities. Topics of interest to the special session include, but are not limited to: Novel deep network architectures for multimodal data representation; Deep learning for new multimedia applications; Efficient training and inference methods for multimedia deep networks; Emerging applications of deep learning in multimedia search, retrieval and management; Deep learning for multimedia content analysis and recommendation; Deep learning for cross-media analysis, knowledge transfer and information sharing; Distributed computing, GPUs and new hardware for deep learning in multimedia research; Other deep learning topics for multimedia computing, involving at least two modalities. Dr. Jinhui Tang, Nanjing University of Science and Technology, China. Dr. Zechao Li, Nanjing University of Science and Technology, China Special Session Web Site

Description: CDeep learning is a topic of broad interest, both to researchers who develop new deep architectures and learning algorithms, as well as to practitioners who apply deep learning models to a wide range of applications, from image classification to video tracking, etc. Brain-like computing combines computational techniques with cognitive ideas, principles and models inspired by the brain for building information systems used in humans’ common life. Pattern recognition is a conventional area of artificial intelligence, which focuses on the recognition of patterns and regularities in data. Recently, there has been very rapid and impressive progress in these three areas, in terms of both theories and applications, but many challenges remain. This workshop aims at bringing together researchers in machine learning and related areas to discuss the utility of deep learning for brain-like computing and pattern recognition, the advances, the challenges we face, and to brainstorm about new solutions and directions. Topics of interest to the special session include, but are not limited to: unsupervised, semisupervised, and supervised deep learning; active learning, transfer learning and multi-task learning; dimensionality reduction, metric learning and kernel learning; sparse modeling; ensemble learning; hierarchical architectures; Doptimization for deep models; intelligent data analysis and recommendation systems; implementation issues, parallelization, software platforms, hardware for deep learning and big data analysis applications in video, image, texture, text processing, neuroscience, medical imaging or any other field. Special Session Web Site

Description: Deep Learning (DL) is growing in popularity because it exploits rather well the unreasonable effectiveness of data to solve complex problems in machine learning. In fact, multi scale machine perception tasks such as object and speech recognitions using DL have recently outperformed systems that have been under development for many years. The principles of DL, and its ability to capture multi scale representations, are very general and the technology can be applied to many other problem domains, which makes it quite attractive. Sponsored by the IEEE Computational Intelligence Society, this event will attract top scientists, researchers, professionals, practitioners and students from around the world. The goal of the IEEE Symposium on DL is to provide a forum for interactions between researchers and practitioners in DL as well as in Artificial Neural Networks, Bayesian Learning, Generative and Predictive Modeling, Optimization, Cognitive Architectures and Machine Learning with an interest in DL. We are interested in discussing the new DL advances, the challenges ahead, and to brainstorm about new solutions and directions. We also seek applications from large engineering firms dedicated to construction and services in energy, autonomous transportation, communications industries, web, marketing, medical and financial services, and scientific fields that require big data analytics. Topics of IEEE DL’17 include but are not limited to: Unsupervised, semi-supervised, and supervised learning Deep reinforcement learning (deep value function estimation, policy learning and stochastic control) Memory Networks and differentiable programming Implementation issues, both software and hardware platforms Applications in vision, audio, speech, natural language processing, robotics, navigation, control, games AI, cognitive architectures, etc. Dimensionality expansion and sparse modeling Learning representations from large-scale data Multi-task learning Learning from multiple modalities Weakly supervised learning Metric learning and kernel learning Hierarchical models Paralleliisation in DL Non-Iterative DL Recursive DL Incremental DL Evolving DL Fast DL Symposium Web Site

Chairs: Nicolò Navarin nnavarin@math.unipd.it, Luca Oneto, Luca Pasa and Alessandro Sperduti. Description: In recent years, Deep Learning has become the go-to solution for a broad range of applications, often outperforming state-of-the-art. Deep learning allows computational models that are composed of multiple processing layers to learn representations of data with multiple levels of abstraction. These methods have dramatically improved the state-of-the-art in speech recognition, computer vision, drug discovery, genomics and many others. Scope and Topics The goal of this special session is to provide a forum for focused discussions on new extensions of deep learning models and techniques, and to gain a deeper understanding of the difficulties and limitations associated with state-of-the-art approaches and algorithms. Practitioners should provide practical insights to the theoreticians, which in turn, should supply theoretical insights and guarantees, further strengthening and sharpening practical intuitions and wisdom. Examples of these possible extensions are: Multimodal and Multitask Deep Learning Deep Transfer Learning Deep Recurrent and Recursive Neural Networks Deep Learning on Structured Data Interpretability of Deep Learning Private and Federated Deep Learning Generative and Adversarial Deep Learning Randomized Deep Learning (Deep ELM, Deep ESN, Deep Reservoir Computing) The focus of this special session is to attract both solid contributions or preliminary results which show the potentiality and the limitations of new ideas, refinements, or contaminations between the different fields of deep learning and other fields of research in solving real world problems. Both theoretical and practical results (e.g. Social Data Analysis, Speech, Natural Language Processing, Cybersecurity) are welcome to our special session. This special session is supported by the IEEE Task Force on Deep Learning . Important dates: Paper Submission Deadline: 15 January 2018 Paper Acceptance Notification Date: 15 March 2018 Final Paper Submission and Early Registration Deadline: 1st May 2018 IEEE WCCI 2018: 08-13 July 2018 Conference Web Site

Chairs: Davide Bacciu (bacciu@di.unipi.it ), Silvio Jamil F. Guimarães and Zenilton K. G. Patrocínio Jr. http://www.icei.pucminas.br/projetos/viplab/ijcnn-deepsm/ A key factor triggering the deep learning revolution has been its ground-breaking performance on image and video processing applications. These have been built mostly on a (multi-dimensional) raw data representation of the visual information. Multimedia content, on the other hand, calls for more articulated data representations catering for the multimodal nature of this information. These are often based on a structured representation that can capture the complexity of the contextual, semantic and geometrical relationships among the visual, phonetic and textual entities and concepts. Scope and Topics: The goal of this special session is to provide a forum for researchers working on the next generation of deep learning models for machine vision and multimedia information, which are capable of extracting and processing information in a structured representation and/or with a multimodal nature. We welcome contributions proposing innovative deep models dealing with: learning hierarchical or networked representations of multimedia information; processing of structured multimedia information under the form of sequences, labelled trees, as well as more general forms of graphs; understanding and synthesizing of multimodal data; fusion of multimodal information. This special session is meant to attract researchers from deep learning, machine vision and multimedia information communities. We aim to bring together researchers with consolidated tradition on structured data processing (such as in machine learning and NLP) with those with machine vision and multimedia processing insight, but mostly working with flat-data representations. Topics of interest for this special session include, but are not limited to, the following: deep learning models for structured data; representation learning in machine vision and multimedia processing; hierarchical/structured visual processing; deep models for visual data streams; generative and variational deep learning for multimedia data; multimedia data synthesis; attentional and bio-inspired models for the processing of visual and audio information; applied deep learning to machine vision and multimedia processing, such as: biomedical images and biobanks, pose and gesture estimation from graphs, etc.; innovative software and libraries for deep learning and multimedia content understanding. Important dates: Paper Submission Deadline: 15 January 2018 Paper Acceptance Notification Date: 15 March 2018 Final Paper Submission and Early Registration Deadline: 1st May 2018 IEEE WCCI 2018: 08-13 July 2018 This special session is supported by the IEEE Task Force on Deep Learning. Conference Web Site

Chairs: Plamen P. Angelov, Lancaster University, UK p.angelov@lancaster.ac.uk Jose C. Principe, University of Florida, principe@cnel.ufl.edu . Synopsis: Deep Learning is becoming a synonym of highly precise (reaching or surpassing capabilities of a human) computational intelligence technique. Very interesting and important results were reported recently in both scientific literature and also grabbed the imagination of the wider public and industry helping propel the interest towards AI, neural networks, machine learning. It was applied mostly to solve classification problems in image processing, but also for predictive tasks in speech processing and other problems. Despite the undoubted success in achieving high precision and avoiding handcrafting in feature selection a number of issues remain unresolved, such as: i) transparency and interpretability; ii) the requirement for extremely large training data set, computational resources and time; iii) overfitting and catastrophic failures with high confidence in some cases; iv) convergence proof for the case of reinforcement learning; v) rigid structure unable to be adapted/to dynamically evolve with new samples and/or new classes; vi) repeatability of the results. Methodologically, the vast majority of the techniques of this hot and quickly developing area are based exclusively on neural networks (convolutional, belief based, etc.). Very recently publications appear where the deep learning (multi-layer) architecture with different levels of abstraction is build based on fuzzy rule-based systems or fuzzy sets are used to represent coefficients/weights in Restricted Bolzman Machines, etc. The aim of the special session is to address the bottleneck issues listed above and discuss and represent alternative and most recent methods, techniques and approaches that can help resolve these issues. The specific sub-topics that will be of interest include: Interpretable/Transparent Deep Learning Computational and time complexity/efficiency of Deep Learning Methods Repeatability of the results of Deep Learning Methods Degree of confidence in the results of Deep Learning Highly Parallelisable Deep Learning Methods Deep Learning with proven convergence Re-trainability and dynamically evolving structures/architectures for Deep Learning Ensembles of Deep Learning Classifiers Fuzzy Deep Rule-based Classifiers Self-adaptive and Self-organising Deep Learning Architectures Also applications to: Computer Vision Image Classification Robotics Remote Sensing Biology and Tomography Surveillance and Defense Industry 4.0 Assistive Technologies and Digital Health Important dates: Paper Submission Deadline: 15 January 2018 Paper Acceptance Notification Date: 15 March 2018 Final Paper Submission and Early Registration Deadline: 1st May 2018 IEEE WCCI 2018: 08-13 July 2018 Submission Guidelines: Please follow the regular submission guidelines of WCCI 2018. Please notify the chairs of your submission by sending an email to: p.angelov@lancaster.ac.uk or principe@cnel.ufl.edu. This special session is supported by the IEEE Task Force on Deep Learning and by Evolving and Adaptive Fuzzy Systems. Conference Web Site