Special sessions

Despite the recent advances in deep machine learning models, there is still a huge gap between machine and human learning processes. Humans can learn from small number of instances and furthermore are able to seamlessly leverage and integrate data from different domains, whereas machines can hardly match this ability even by learning from big data. Following the way humans process data and learn from it, recent years have witnessed the development of models to imitate human-like learning paradigm. In particular domain adaptation, transfer learning, zero-shot, one-shot or few-shot learning, incremental learning, active learning, multi-task and multi-view learning are examples of such models. One can anticipate that the future methodologies should be much more flexible for learning in many application domains including biomedical signal and medical image analysis, image/video classification, localization and segmentation, information retrieval and complex systems among others. Therefore, the main objective of the session is to discuss the recent rise of new research questions and learning strategies for the following problems using both shallow and deep models. The topics of interest include but not limited to:

• Deep and shallow models
• Neural Networks/ Kernel models
• Transfer learning 
• Domain adaptation 
• Zero /One / Few-shot learning
• Feature learning / Representation learning 
• Supervised /  Semi-supervised / Unsupervised learning
• Multi-task learning
• Multi-view learning
• Multi-label learning 
• Learning from heterogeneous data

The application of machine learning techniques in the area of computer networks is a very promising area of study. Steadily growing network traffic, especially in the Internet or in huge data centers, as well as the continuously increasing number of end points due to mobile devices and IoT applications, implies an enormous management and monitoring effort when designing and operating these networks. Network automation mechanisms based on, e.g., software-defined Networking (SDN), can leverage advanced telemetry solutions to allow fine-grained traffic management. Large scale data transfer and ever-growing bandwidths raise the demand for open-loop network management assistance or sustained closed-loop auto-remediation, self-healing or -optimization techniques.

On the machine learning side, challenges arise due to the inherent "big data" aspect of network traffic. Because of this, learning has to be conducted "on the fly" on streaming data, without storing any data at all. Learning is further complicated by the non-stationarity of the data which can produce the catastrophic forgetting effect to which DNNs are particularly vulnerable. Last but not least, the acquisition of sufficient amounts of good-quality training data is often difficult due to privacy protection issues, and the results of machine learning are often not generalizable because all networks and their users have strongly individual characteristics.

We encourage submission of papers on novel methods or applications of machine learning for computer network management and monitoring, including but not limited to:

• machine learning for network automation and programmability in the data, control, management or knowledge plane cognitive/autonomic network management and monitoring
• machine learning in fault, configuration, accounting, performance or security management
• big data and deep Learning Approaches for network traffic engineering and routing
• in-network computing based on machine learning
• streaming (network) data processing by deep neural networks
• continual learning on network traffic data
• acquisition of training data from computer networks and generalization of results

In the latter years, tensor decompositions have been gaining increasing interest in the Big Data and Machine Learning community.  On the one hand, multiway data analysis provides powerful and efficient methods to address the processing of large scale and highly complex data, such as multivariate sensor signals.  On the other hand, tensor decompositions have been shown to provide the necessary theoretical backbone to study the expressiveness properties of deep neural networks. More recently, tensors have started to find wide application in a variety of machine learning paradigms, ranging from neural networks to probabilistic models, to enable the efficient compression of the model parameters leveraging a wide range of decomposition methods from multilinear algebra.

This special session aims to present the state­of­the­art on this increasingly relevant topic among ML theoretician and practitioners. To this end, we welcome both solid contributions and preliminary relevant results showing potential, limitations and challenges of new ideas related to the use of tensor decompositions in deep learning, neural networks, and machine learning at large.  Studies stemming from major research initiatives and projects focusing on the session topics are particularly welcome. Topics of interest include theory and applications of tensors in deep neural networks, including but not limited to:

• Tensor decompositions and multiway data analytics
• Tensor decompositions for signal processing
• Tensor neural networks
• Tensor-based representation and manipulation of neural weights
• Decomposing neural architectures through tensor manipulation
• Theoretical analysis of neural networks through multiway algebra
• Use of tensor decompositions in learning systems
• Speeding up neural computations through tensor decompositions
• Tensor based approaches for structured data processing (graphs, trees, sequences)
• Tensors and dynamic memory (recurrent neural networks)
• Applications of tensor neural networks to image and video processing
• Applications of tensor neural networks to sensor and stream data analysis

Natural Language Processing (NLP) is a branch of artificial intelligence brimful of intricate, sophisticated, and challenging tasks, such as machine translation, question answering, summarization, and so on.

Thanks to the recent advances of deep learning, NLP applications have received an unprecedented boost in performance, generating growing interest from the machine learning community. However, even if deep learning techniques are starting to reach excellent performance on various tasks, there are still several problems that need to be solved, such as the computational cost, the reproducibility of results, and the lack of interpretability

This special session is devoted to cover novel research about popular and fashionable applications of machine learning and deep learning solutions to NLP tasks. Both theoretical analysis and empirical solutions to real-world problems are welcome and encouraged.

Relevant deep learning topics include, but are not limited to:

• word, sentence, and document embeddings
• graph embeddings and knowledge bases
• information extraction
• named entity recognition and relation extraction
• question answering
• string kernels
• encoding/decoding architectures
• summarization
• machine translation
• document classification
• sentiment analysis
• efficient and interpretable models
• ontologies and knowledge-driven approaches
• resources, datasets, and pre-trained models
• multidisciplinary applications (such as biomedical text mining, verbal lie detection...)
• creative tasks (such as storytelling, caption generation, social media analytics, multilinguality…)

Modern high-throughput technologies deployed in research and development for new pharmaceutical products have opened the door to machine learning applications that allow the automation of tasks and support for data-driven risk-based decision making. Furthermore, the development of modern complex drugs requires more advanced technologies during the exploration of a molecule candidate than those used in previous generations.  These new technologies typically generate a much larger amount and variety of data and therefore require much more advanced data integration and analysis techniques. Down the road, biomanufacturing production lines generate heterogeneous data in almost real-time, which can also be used to optimize yield or quality, as well as prevent unexpected down-times.

Despite the numerous appealing opportunities, many practical challenges make it difficult to apply machine learning models in the pharmaceutical field, such as the large intrinsic variability of the process, the large imbalance in labels, high dimensionality with respect to the number of observations, noisy data collection, the large variety of data types (multi-omics, images, time series, spectra, real-world evidence etc.), regulatory considerations, online batch production data, etc. For these reasons, very innovative algorithmic and mathematical approaches are often needed.

We encourage submission of papers on novel machine learning, computer vision and natural language approaches applied to drug discovery, production and quality control in the pharmaceutical industry, including but not limited to

• Drug design, protein folding etc.
• Next-generation sequencing and multi-omics analysis
• Biomedical image and text analysis
• Learning from heterogeneous data types 
• Algorithms for unbalanced datasets
• Online quality control with process analytical technology
• Data drift and shift handling, transfer learning
• Learning from noisy labels and weakly annotated data
• Low shot and transfer learning
• Active learning
• Transparency, security and privacy of models from patient data
• Active learning and novel approaches to labelled data set creation 

Machine Learning (ML) is a very prolific field of research. As a result of this, new paradigms and novel contributions have extensively been proposed in the last few years. One of the most disruptive ones is undoubtedly Quantum Machine Learning (QML). QML combines concepts from ML and Quantum Physics (QP), particularly Quantum Computing (QC), to provide enhanced alternative solutions to problems of data analysis.

QML has already shown promising ways to speed up some of the costly ML calculations without worsening the performance. However, this view is somewhat limited; QML should not only be seen as a method that makes use of QC or QP to improve classical ML but it can also encompass problems in which ML - either classical or quantum – can provide efficient solutions to problems of QP that are especially difficult to be modeled.

The community working on QML mostly comes from the field of Physics but the contribution of ML researchers is also crucial to come up with sound solutions from the point of view of both disciplines. Therefore, it is very relevant to have the chance to meet up with other colleagues from both fields in order to be aware of the last contributions and discuss current ideas. This special session is hence aimed at providing a discussion forum for researchers working on QML, and ML applications to Physics.

The main topics of the session include, but are not limited to, theoretical developments in, and applications of:

• Adiabatic quantum optimization
• Machine learning for scientific and technological applications
• Quantum biomimetics
• Quantum clustering
• Quantum computing
• Quantum information
• Quantum learning
• Quantum neural networks
• Quantum principal component analysis
• Quantum regression
• Quantum reinforcement learning
• Quantum support vector machines
• Quantum technologies

Reservoir computing (RC) studies the properties of large recurrent networks of artificial neurons, with either fixed or random connectivity. Over the last years, reservoirs have become a key tool for pattern recognition and neuroscience problems, being able to develop a rich representation of the temporal information even if left untrained. 
The common paradigm has been instantiated into several models, among which the Echo State Network and the Liquid State Machine represent the most widely known ones.
Nowadays, RC represents the de facto state-of-the-art approach for efficient learning in the temporal domain. Besides, theoretical studies in RC area can contribute to the broader field of Recurrent Neural Networks research by enabling a deeper understanding of the fundamental capabilities of dynamical recurrent models, even in the absence of training of the recurrent connections. RC paradigm also allows using different dynamical systems, including hardware, for computation.

This special session targets the latest contributions emerging in the field of RC from all points of view, including advancements in theory, applications, and hardware implementations. Comparisons with traditional and deep learning models, analyzing the impact of the topology of the reservoir are also welcomed. Overall, the special session aims at stimulating an open discussion in the neural network community, pushing to a broader understanding of RC models and neural networks in general.

Topics of the special session:

• Novel Reservoir Computing models
• Deep Reservoir Computing
• Reservoir Computing for Big Data
• Physical, Neuromorphic and Photonic implementations of Reservoir Computing
• Theoretical analysis of Reservoir Computing models (e.g., approximation abilities, asymptotic properties, etc.)
• Echo State Property and stability of input-driven reservoirs
• Architectural design of reservoirs
• Evolutions of the RC paradigm (e.g., conceptors)
• Biological motivations and applications in neuroscience
• Adaptation of reservoir dynamics and of system dynamics
• Non-iterative methods for readout training
• Novel application fields for the RC paradigm