Generically trained SFA networks for cognitive mapping in simulated rodents Computational Neuroscience


Slow feature analysis (SFA) techniques have been used to simulate the self-organization of spatially tuned neural activity in synthetic rodents, see Franzius et al. (2007), and Schönfeld et al. (2013, 2015). We have implemented these approaches in a virtual reality (VR) environment, driven by the Blender game engine. While Blender allows for the swift creation of novel environments, and straightforward environmental context manipulation, the employed SFA networks require an extensive training phase in each new surrounding. This is not only problematic from the computational point of view, but is also inconsistent with biological observations: rodents seem to build representations of new environments quickly, without the need for extensive training. To address this issue, the use of generically trained SFA networks in different environments will be assessed within the project’s context. Adequate training data might be generated through the Blender-based VR system, or might be sampled from real-world footage. The generically trained, hierarchical SFA networks will be ‘implanted’ into a virtual robot that is then subjected to varying environmental contexts. We expect evolution of generic cognitive maps that can be used to bias further spatial neural tuning of place cells in the robot model. The robustness of these maps will be assessed in a reasonably large range of different scenarios. We also assume that generic SFA sub-networks that were trained on simple environmental features (e.g., edges, crossings) can be assembled in a superordinate circuit that either renders environment-specific training unnecessary, or at least causes a significant speed-up in the acquisition of cognitive maps in new scenarios. To master the above challenges, an excellent command of Python (including numpy, matplotlib, scipy) and C/C++ is required. Experience with the Blender simulation environment would be advantageous.


Franzius, M., Sprekeler, H., & Wiskott, L. (2007). Slowness and sparseness lead to place, head-direction, and spatial-view cells. PLoS Computational Biology, 3(8), e166.

Schönfeld, F., & Wiskott, L. (2013). RatLab: an easy to use tool for place code simulations. Frontiers in Computational Neuroscience, 7, 104.

Schönfeld, F., & Wiskott, L. (2015). Modeling place field activity with hierarchical slow feature analysis. Frontiers in Computational Neuroscience, 9, 51.


Dr.-Ing. Thomas Walther and Prof. Dr. Sen Cheng, Institute for Neural Computation

The Institut für Neuroinformatik (INI) is a central research unit of the Ruhr-Universität Bochum. We aim to understand the fundamental principles through which organisms generate behavior and cognition while linked to their environments through sensory systems and while acting in those environments through effector systems. Inspired by our insights into such natural cognitive systems, we seek new solutions to problems of information processing in artificial cognitive systems. We draw from a variety of disciplines that include experimental approaches from psychology and neurophysiology as well as theoretical approaches from physics, mathematics, electrical engineering and applied computer science, in particular machine learning, artificial intelligence, and computer vision.

Universitätsstr. 150, Building NB, Room 3/32
D-44801 Bochum, Germany

Tel: (+49) 234 32-28967
Fax: (+49) 234 32-14210