Computing and Learning in Circuits of Spiking Neurons

Information is processed in the neocortex by extremely complex but surprisingly stereotypic circuits of neurons. The goal of this project is to understand the organization of computation and learning in neural microcircuits, which form the lowest level of circuit organization in the cortex. This research will be carried out through theoretical analysis and computer simulation, which take the most recent biological data into account. A unique aspect of this project is the close collaboration of computer scientists and neuroscientists. The team of Prof. Maass at the Institute for Theoretical Computer Science (TU Graz) will collaborate with the team of the neuroscientist Prof. Markram from the Weizmann Institute in Israel, who is one of the leading experts for the experimental investigation of neural microcircuits in the neocortex. The project builds on results from a preceding FWF-project, that had focused on the components of neural circuits: spiking neurons and dynamic synapses. It is expected that the understanding of the organization of information processing in neural circuits will also provide new ideas for the design of novel artificial computing machinery ("neuromorphic engineering"). In this project computer models of neural microcircuits will be used to explore new methods for training a robot to respond in real-time to rapidly changing input.

Most preceding models for the computational function of millions of local circuits of neurons in the cerebral cortex, so-called cortical microcircuits, had focused on the assumption that these circuits have genetically encoded fixed computational functions, although there is no convincing experimental evidence for this assumption. In contrast, the work of this project has explored the hypothesis that generic cortical microcircuits have a more general computational task, such as temporal integration of information and nonlinear combination of input variables (similar as a kernel in machine learning). In this way such cortical microcircuit could support (as a common pre-processor) simultaneously the computational task of thousands of different readout neurons, that extract information from a local microcircuit, and project specific aspects and combinations of this information to other circuits, cortical areas, or to subcortical structures. Our theory proposes that the specific role of each such readout neuron is shaped through learning, and we have shown that an experimentally supported learning rule for spike-timing-dependent plasticity (STDP) would in principle suffice for that. We have demonstrated in this project through mathematical analysis and extensive computer simulations, that his theory is consistent with numerous known details of cortical microcircuits, and that the resulting computational power of cortical microcircuits is unexpectedly large. Neurobiological experiments are currently carried out, which test specific predictions of this new model. Of particular interest from a more general perspective is that our model for computations in cortical microcircuits is suitable for implementing "anytime algorithms". This means, that one does not have to wait until the algorithm provides a result. Rather, estimates of this result become available at any time during the computation, and are automatically improved if one can afford to wait longer. The work of this project, together with the simultaneous work of Herbert Jaeger at the International University of Bremen, has provided the first successful methods for designing circuits for anytime computing. The project has demonstrated the technological relevance of these methods through applications in robotics and computer vision. Researchers in other countries have subsequently found other interesting applications.