Special sessions

Machine learning approaches and artificial intelligence techniques contribute more and more to advances in molecular medicine and biology, as well as in other branches of bio-medical research, which all have to deal with increasing amounts of data to be processed and analyzed.

Respective methods and concepts are required to be robust, transparent, and information-preserving. Moreover, it is beneficial for the domain experts if knowledge acquired by the model for decision making is provided to them as users. Thus, interpretable or at least explainable models are desirable and favored in this area although they are not frequently available yet.

The aim of the session is to present respective machine learning approaches with the application in molecular biology and medicine. Possible application topics could be

• Sequence analysis
• Medical diagnosis support systems
• Image analysis and understanding based on machine learning and deep learning
• Gene expression analysis
• Intelligent data visualization
• ...

For the machine learning approaches we encourage authors to provide contributions regarding

• Interpretable and explainable methods
• Information-theoretic models and information extraction
• Knowledge representation
• Robust methods and model confidence, certainty of predictions

Yet other approaches and topics orbiting around the main focus are welcome. Further, we also appreciate contributions dedicated to methodological aspects and new ML-paradigms for bio-medical data analysis.
 

Continual Learning is a fast-growing recent field of research which deals with learning from non-stationary data distributions. Since this violates a key premise of statistical learning theory, common ML models such as DNNs/CNNs, in their original form, are unsuited for CL and often suffer from an effect known as "catastrophic forgetting".

CL is mainly studied for classification tasks, e.g., sequentially training MNIST classes in disjoint groups. The objective of this special session is to study CL for flavors of machine learning other than classification.

Typical examples are unsupervised learning (e.g., GMMs, SOMs), generative models (typically GANs, VAEs, flow models) or reinforcement learning. Since the assumptions about the nature of the non-stationarity are usually very different from those made for classification problems, we expect that different approaches will be required as well, thus significantly expanding our current view of CL.
 

Recently, deep learning models have had a huge impact on computer vision applications,
in particular in semantic segmentation, in which many challenges are open. As an
example, the lack of large annotated datasets implies the need for new semi-supervised
and unsupervised techniques. This problem is particularly relevant in the medical field
due to privacy issues and high costs of image tagging by medical experts. The aim of this
session is to provide an international forum for presenting recent results and advances
regarding machine learning methods in image processing and computer vision.

The scope of this session comprehends, but it is not limited to:

• biomedical image processing
• image forensic
• anomaly detection
• 3d image segmentation

Contemporary data science applications are continuously being challenged by the complexity and the heterogeneity of the collected data. Whether we speak of vector data, or of text, graph, and other (semi)structured data, identifying abnormal records is one of the first specific challenges for the analysis. Furthermore, since the data is usually collected with a temporal and/or a spatial stamp, the issue of partitioning it into homogeneous segments and detecting change-points is another major challenge, strongly related to the first one.

Numerous research directions have been explored in this context, including, among others the following ones:

• vector embedding
• dissimilarity based methods
• generative models
• kernel methods
• deep learning
• etc.

This special session is thus driven by two related and complementary topics: on the one hand, outlier detection, and on the other, anomaly and change-point detection in data with a temporal and/or a spatial dimension, whether in online and/or offline fashion.

We encourage submission of theoretical results, methodological advances and novel applications. A particular interest will be shown to proposals focusing on non-vector data, such as categorical data (especially when the number of categories is large or when some structure exists over the categories), graphs (including temporal, attributed, etc.), texts (including collections or evolving through time), and other (semi)structured data (XML documents, computer programs, etc.).

Traditional deep learning approaches have been developed assuming data to be encoded into feature vectors, however many important real-world applications generate data that are naturally represented by more complex structures, such as graphs. Graphs are particularly suited to represent relations between the components constituting an entity, allowing us to effectively describe systems of interacting elements, like social, biological, and technological networks, as well as data where topological variations influence the feature of interest, e.g., the interaction of proteins or molecular compounds.

This has motivated a recent increasing interest of the machine learning community in the development of learning models for structured information.

The field of graph deep learning, in particular, combines the ability of deep neural networks to learn representations end-to-end with this explicit description of relations in the data. Specifically, the class of models at the heart of graph deep learning, generically called Graph Neural Networks (GNNs), extend and generalize typical convolutional neural networks to process arbitrary graphs.

Topics of interest to this session include, but are not limited to:

•     Graph Neural Networks: theory and applications
•     Graph representational learning
•     Graph generation (probabilistic models, variational autoencoders, adversarial learning, etc.)
•     Graph learning and relational inference
•     Graph kernels and distances
•     Scalability, data efficiency, and training techniques of graph neural networks
•     Deep learning for dynamic graphs and graph sequences
•     Reservoir computing and randomized neural networks for graphs
•     Recurrent, recursive and contextual models
•     Graph datasets and benchmarks
•     Applications in natural language processing, computer vision (e.g. point clouds), materials science, cheminformatics, computational biology, social networks, etc.