Structural plasticity and multi-time-scale learning in physical reservoirs



Dr. Hans Kleemann

ODS (Organic Devices and Structures) working group, Chair of Optoelectronics, TU Dresden


Prof. Dr. Christian Tetzlaff

Institute for Neuro- and Sensory Physiology Department of Computational Synaptic Physiology, University Medical Center Göttingen

Project Description

The human brain can efficiently learn to solve complex and versatile tasks, and it continues improving its performance throughout our biological development from the age of a child to an adult. These abilities are based on the diverse mechanisms of synaptic plasticity, which allow synapses to store and process information over a broad time scale ranging from milliseconds to weeks. Furthermore, structural plasticity, i.e. growth and depletion of synaptic connections, which takes place throughout our life span, plays a crucial role in the development of efficient learning rules in the brain. So far, this biological multi-time-scale and structural plasticity can only be realized to a very limited extent or not at all in electronic systems for neuromorphic computing. On the one hand, analog components with simultaneously volatile and non-volatile memory functions are missing, and on the other hand, there is also no possibility of physically creating or removing connections on a CMOS chip during learning.

In this project, we develop experimental and theoretical methods to realize and study synapses and recurrent neural networks (so-called reservoirs) with multi-time-scale and structural plasticity. We use polymer-based dendrites whose synaptic properties can be dynamically controlled during growth, and we analyze the interaction of different plasticity learning rules, spanning over the biologically relevant time scale from milliseconds to weeks in these systems. In this context, we investigate how the chemical composition of the polymers and electrolytes can be utilized to tune the strength of the different plasticity mechanisms. Furthermore, we analyze the transition from the single synapse to recurrent neural networks in experiment and theory, and based on the dynamics and storage capacity of the system, we define learning rules that can be used, for example, for time series prediction.

Ultimately, due to the biocompatibility of the selected substrates and materials, our system is ideally suited for application in medical and biological electronic devices. Therefore, we demonstrate in experiment and theory the performance of the reservoirs and the efficiency of the learning rules for the prediction of biosignals, which have broad temporal dynamics, such as brain activity and body temperature.

Further involved scientists


Dr. Erika Covi

Cognitive Devices Group at Groningen Cognitive Systems and Materials Center (CogniGron), University of Groningen