SFB 1597:  Small Data

• Subject Area: Medicine ; Biology ; Computer Science, Systems and Electrical Engineering ; Mathematics ; Physics ; Social and Behavioural Sciences
• Term: since 2023
• Website: Homepage
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 499552394

The recent progress in artificial intelligence has been facilitated by big data volumes and data-driven modeling approaches. In particular, this includes deep learning techniques, mainly developed in computer science. However, there is a much larger number of applications, particularly in biomedical settings, where data analysis has to be performed with a relatively small number of observations. In data-driven modeling, this can be addressed by transferring additional information. Alternatively, modeling can be based on imposing stronger structural assumptions, e.g., reflecting the input of biomedical experts, in knowledge-driven approaches, frequently developed in mathematics and statistics/systems modeling. Thus, approaches for small data challenges are currently scattered across different data science disciplines. For creating comprehensive solutions that fuse exciting emerging ideas, it is therefore necessary to integrate contributions from computer science, mathematics, and statistics/systems modeling. This methods development also needs input from application domains, such as biomedicine. Correspondingly, we have designed our CRC Initiative SmallData with a strong focus on developing an interdisciplinary methods framework. We focus on the key small data tasks of combining similar datasets and transferring information from additional sources, while taking into account and reducing uncertainty. This is reflected in the three key areas of SmallData, similarity, transfer, and uncertainty. In terms of methods, we focus on combining data-driven and knowledge-driven modeling approaches, e.g., based on neural networks and differential equations. Meta-learning and pre-training are two further components of our framework for transferring information on model parameters or tuning parameters between datasets. We have furthermore designed a Fusion Hub, which addresses overarching topics that link the three areas, by fusing methods from different disciplines. This includes approaches based on the concepts of attention and few-shot learning, which have recently been advanced in computer science. Development of theory will go hand in hand with tailoring methods to prototypical biomedical applications from forensic medicine, gene therapy, nephrology, psychiatry, radiology, and rare diseases. In addition, our integrated research training group will foster a shared language across disciplines. We have furthermore designed the SmallData Compendium as a web platform that will reflect our interdisciplinary methods framework for exchanging concepts and methods with the international community, and thus shaping the small data field.

• DFG Programme: Collaborative Research Centres

Current projects

• A01 - Integrating information from similar data sites when developing clinical prediction models for patients from a target site (Project Heads Rohde, Angelika ;  Zöller, Daniela )
• A02 - Identifying best practice treatment strategies by incorporating information from similar healthcare pathways (Project Heads Binder, Nadine ;  Dette, Holger )
• A03 - Similarity of individual latent dynamics in longitudinal cohort data (Project Heads Binder, Harald ;  Schmidt, Thorsten )
• A04 - An effective similarity integration multi-modal graph neural network method to facilitate disease gene prioritization (Project Heads Backofen, Rolf ;  Schmidts, Miriam )
• A05 - Enabling efficient and safe application of CRISPR-Cas in primary human cells by deep learning-based information transfer from well-investigated cell types (Project Heads Backofen, Rolf ;  Cathomen, Toni )
• B01 - Transfer- and meta-learning in deep networks for human brain-signal analysis (Project Heads Ball, Tonio ;  Hutter, Ph.D., Frank )
• B02 - Transfer learning for forecasting short environmental time series using process- guided neural networks (Project Heads Bödecker, Joschka ;  Dormann, Carsten )
• B03 - Linking cohorts and expert knowledge through categorical representations for improved knowledge extraction from longitudinal data (Project Heads Has, Cristina ;  Hess, Moritz )
• B04 - Essentials for few-shot learning on images (Project Heads Brox, Thomas ;  Valada, Abhinav )
• B05 - Cross-modal representation learning, with applications to search in radiology reports and auto-filling of report templates (Project Heads Bast, Hannah ;  Brox, Thomas ;  Kotter, Elmar )
• B06 - Reciprocal transfer of knowledge from population-based genetic screens to whole-body, organ-resolved models of human metabolism (Project Heads Hertel, Johannes ;  Köttgen, Anna )
• C01 - Uncertainty quantification in classification with applications in forensic genetics (Project Heads Lutz-Bonengel, Sabine ;  Pfaffelhuber, Peter ;  Rohde, Angelika )
• C02 - Uncertainty and heterogeneity in network meta-analysis with small subgroups and few studies (Project Heads Nikolakopoulou, Adriani ;  Schramm, Ph.D., Elisabeth )
• C03 - Reducing parameter optimization uncertainties of dynamic models by meta- learning (Project Heads Kreutz, Clemens ;  Timmer, Jens )
• C04 - Learning fast and efficient hyperparameter control for deep reinforcement learning on small datasets (Project Heads Awad, Ph.D., Noor ;  Bödecker, Joschka )
• C05 - Meta-learning for regularizing deep networks under small data regimes (Project Heads Grabocka, Josif ;  Hutter, Ph.D., Frank )
• F - Fusion Hub (Project Head Binder, Harald )
• MGK - Integrated research training group (Project Heads Backofen, Rolf ;  Binder, Nadine )
• Z - Central administrative project (Project Head Binder, Harald )

• Applicant Institution: Albert-Ludwigs-Universität Freiburg
• Participating University: Ruhr-Universität Bochum ; Fakultät für Mathematik; Universitätsmedizin Greifswald; Klinik und Poliklinik für Psychiatrie und Psychotherapie
• Spokesperson: Professor Dr. Harald Binder