Special Sessions

  1. BIGCHEM Special Session
  2. Artificial Intelligence in Medicine
  3. Informed and Explainable Methods for Machine Learning
  4. Deep Learning in Image Reconstruction
  5. Machine Learning with Graphs: Algorithms and Applications
  6. Neural Variational Inference and Time Series

Note: submissions to the Special Sessions follow the same procedure and deadlines of the conference (see the Submission page for details).

Contents and organisers

1. BIGCHEM Special Session

Artificial Intelligence and machine learning are increasingly used in the chemical industry, in particular with respect to Big Data. These developments have the potential to automate, facilitate and speed-up the key steps in drug research, however their applications are still at an early stage. In particular, this is due to the need to develop “chemistry aware” methods and/or adapt existing methods to work with chemical data. The goal of this session is to show progress and exemplify the current needs, trends and requirements for AI and machine learning for chemical data analysis. In particular it will focus on the use of chemical informatics and machine learning methodologies to analyse chemical Big Data, e.g., to predict biological activities and physico-chemical properties, facilitate property-oriented data mining, predict biological targets for compounds on a large scale, design new chemical compounds, and analyse large virtual chemical spaces.

ICANN2019 and Springer Open Access collaboration. The authors of articles/abstracts submitted to the BIGCHEM Special Session are qualified for a 25% discount on the journal’s article-processing charge for the special issue of J. Cheminformatics. To be qualified for the discount one of the authors should participate to the ICANN2019 and the article should be submitted to the special issue before the start of the conference. The article submitted to J. Cheminformatics will be fully peer-reviewed and should comply to the usual publishing ethics.

Instructions: To be qualified for the discount one of the authors should submit an abstract (or preliminary/related study) to the conference proceedings to be published by Springer in Lecture Notes in Computer Science, register and participate to the ICANN2019. The journal article should be submitted to the special issue before the start of the conference. 


  • Big Data analysis in chemistry
  • Machine Learning in drug-discovery
  • Machine learning methodologies for mining very large compound data sets
  • Big data visualization and modelling
  • Virtual screening methods to exploit large virtual chemical spaces
  • Machine learning and the use of HTS data for compound activity predictions
  • Analysis of compound promiscuity and frequent hitters
  • Reaction-driven de novo design to explore the chemical space
  • Accessing new chemical space based on predictive models
  • Reaction informatics and synthetic route prediction
  • Federated learning and secure sharing of information
  • Molecular dynamics and quantum chemistry calculations using neural networks

Organizers: BIGCHEM Consortium

2. Artificial Intelligence in Medicine

In conventional settings, diagnosis of diseases and further medical and surgical treatment choices are mostly subjective, based on limited assessment of patient phenotypes. In the era of petabytes of clinical data and detailed patient information, from genomics to histological or volumetric images, over the span of entire lives, AI is instrumental to assess the decisions made for patients through the lens of understanding patient treatment via big data analysis.
In this session, we discuss state-of-the-art AI methods available to analyze patient multimodal data, understand patients’ response to medication and
treatments, as well as to predict risk of diseases prior to onset, which would
potentially provide time for better treatment plans.
Areas of Interest: We invite all researchers in the field to participate in our session, by submitting their original work and/or registering to participate in the following topics:

  1. Methods for clinical diagnosis and outcome prediction
  2. Identification of increased-risk patient categories
  3. Ethics and security of healthcare data used by AI
  4. Methods for computational pathology


  1. Dr. Carsten Marr, Helmholtz Zentrum München
  2. Dr. Narges Ahmidi, Helmholtz Zentrum München and Johns Hopkins University
  3. Dr. Tingying Peng, Helmholtz Zentrum München

3. Informed and Explainable Methods for Machine Learning

Although the latest advancements in deep learning have pushed the boundaries of artificial intelligence and already resulted in successful solutions for very challenging problems, two major practical aspects, which are of utmost importance in a variety of industry applications, remain mostly uncovered. The first aspect is related to lack of training data (i.e. having thin data scenarios), which due to the Vapnik-Chervonenkis theory is a crucial issue to train well-generalizing deep models with millions of adjustable parameters. Whereas, the second aspect is related to the lack of traceability as the typical connectionist models can be seen as black boxes since their inner computations become difficult-to-interpret with the increased complexity. In this session we aim to bring together lead researchers from industry and research to concentrate on methods to incorporate knowledge into the state-of-the-art machine learning models to reduce the need for massive training datasets and investigate a variety of methods to obtain interpretable solutions.

The main list of non-comprehensive topics of this special session will be:

  • Interpretability in deep neural networks
  • Monte Carlo tree search
  • Interpretable matrix and tensor factorization models
  • Bayesian methods for interpretability
  • Interactive and online machine learning
  • Informed reinforcement learning
  • Hybrid and fuzzy AI solutions
  • Expert Systems
  • Self-assessment for supervised and unsupervised learning methods
  • Visualization of machine learning methods
  • Practical applications with informed and/or interpretable machine learning methods


  1. Prof. Dr. Christian Bauckhage; Head of Fraunhofer Center for Machine Learning and Professor of Computer Science in University of Bonn.
  2. Rafet Sifa; Lead Data Scientist and Head of Cognitive Business Optimization group at Fraunhofer IAIS.

4. Deep Learning in Image Reconstruction

Background: Deep learning (DL) has recently gained a lot of attention as superior method in solving challenging problems in machine learning. Many computer vision problems have been successfully tackled by DL algorithms. Recently, researchers have started to apply DL to image reconstruction for various modalities. A bottleneck in this approach arises from the lack of labeled data for training large-scale networks. To overcome this, simulations of the forward model for a given imaging modality have been used to create sufficient data for training. The reconstruction problem can then be formulated as a supervised learning problem. Taking the forward computation into account, such approaches typically result in a kind of autoencoder for solving the inverse problem in image reconstruction, thus incorporating additional knowledge in form of the forward model into the data-driven approach.

This special session aims at bringing together researchers working in this field. By discussing the trade-offs between the data-driven approach of DL and the traditional model-driven reconstruction algorithms, we hope that we will improve our understanding of strengths and weaknesses of DL algorithms in image reconstruction. Overall, making better use of the underlying model of the imaging modality at hand is expected to lead to smarter algorithms for data-driven image reconstruction.


  • Werner Dubitzky, Helmholtz Zentrum München
  • Keiichi Ito, Helmholtz Zentrum München
  • Carlos Garcia Perez, Helmholtz Zentrum München
  • Wolfgang zu Castell, Helmholtz Zentrum München

5. Machine Learning with Graphs: Algorithms and Applications

Networks have become ubiquitous as data from many diverse disciplines can naturally be modeled as graph structures. Characteristic examples include social and information networks, technological networks, web graphs, as well as networks from the domain of biology and neuroscience. Developing machine learning algorithms for graph data is a crucial task with a plethora of cross-disciplinary applications. The special session aims to bring together researchers from both academia and industry that are interested in state-of-the-art algorithmic techniques and methodologies in machine learning for graphs and their related applications.
Topics of interest include, but are not limited to:

  • Algorithms and methods
    • Representation learning on graphs
    • Deep learning and graph neural networks
    • Probabilistic graphical models
    • Statistical models of graphs
    • Graph kernels and graph similarity
    • Semi-supervised graph learning
    • Graph sampling and inference
    • Graph clustering and community detection
    • Graph summarization
    • Anomaly detection in networks
    • Mining and learning from temporal and dynamic networks
    • Learning on heterogeneous graphs
  • Application domains
    • Social media and social network analysis
    • Knowledge graphs and semantic networks
    • Communication networks
    • Recommender systems
    • Urban network analysis
    • Analysis of biological and ecological networks
    • Neuroscience and brain network analysis


  • Danai Koutra (University of Michigan, Ann Arbor)
  • Fragkiskos Malliaros (CentraleSupélec, University of Paris-Saclay and Inria Saclay)
  • Evangelos Papalexakis (University of California Riverside)

6. Neural Variational Inference and Time Series

Variational Auto-Encoders (VAEs) have proven to be a powerful model for learning generative models of complex data with widespread applications. Leveraging the flexibility of neural networks, VAEs are able to model and learn generative models of challenging data, together with suitable approximate inference networks. The prospect of powerful generative models has sparked a wide array of research into neural latent-variable models and variational inference for time series. This has proven to be an interesting topic: generative models for time series allow for more assumptions and inductive biases, e.g. state-space models or auto-regressive models. At the same time, posterior distributions in time-series models have much richer structure than for static data.
Researchers are slowly beginning to understand the trade-offs for different models in terms of generative fidelity, likelihood maximization, and reliable inference.
This special session aims at bringing together researchers working on neural latent variable models and their applications, as well as researchers from neighbouring fields like dynamical systems, reinforcement learning, robotics, and others.
Topics of interest include, but are not limited to:

  • Latent-variable models for time series
  • Neural networks and graphical models
  • Approximate inference in time series
  • Neural filtering, smoothing, and prediction
  • Learning and identifying dynamical systems
  • Applications of (neural) variational inference to time series data, including but not limited to:
    • Reinforcement learning
    • Control
    • Robotics
    • SLAM


  • Maximilian Sölch, Volkswagen Machine Learning Research Lab, Munich
  • Justin Bayer, Volkswagen Machine Learning Research Lab, Munich
  • Patrick van der Smagt, Volkswagen Machine Learning Research Lab, Munich