| dc.description.abstracteng | It is still a mystery how information is processed in the brain, dynamically and reliably at the same time, in particular when considering that a central nervous system operates mainly with information that is processed locally at and between single neurons. To analyze how movement could be represented in the brain, I aimed to find new approaches for solving basic neural encoding problems with local-only mechanisms, performed on the neuromorphic hardware Loihi. The theoretical hierarchical motor selection and execution (HMSE) model suggests biologically plausible mechanisms for movement encoding. A crucial issue for movement execution is the so-called stability-variability trade-off, the problem of obtaining enough variability in neural activity to provide sufficient information for learning complex movement trajectories, while keeping a sufficient level of robustness against noise at the same time. To achieve this, I analysed and successfully applied a new type of network which is based on an inhomogeneous connectivity structure, called anisotropic network. But even with stable and variable data at hand, it is still unclear how this information can be used to represent low-dimensional trajectories, like movements. Thus, I applied a reward-based supervised output learning mechanism, premised on the ReSuMe learning rule, that uses purely local synaptic mechanisms and allows learning of functions based on the spiking behavior of a network. In addition, most simulations were performed using the neuromorphic hardware Loihi. In order to enable the above mentioned simulations on this specialized hardware, I developed a variety of open source tools supporting neuroscientific experiments on Loihi, including a reservoir computing framework, an emulator for faster prototyping and a tool for translating parameters of existing neuron models to the Loihi neuron model. With these tools and frameworks at hand, I demonstrated that the neuromorphic chip Loihi can successfully be used for neuroscientific research. Overall, the results provide new insights into possible encoding schemes of biologically plausible spiking neural networks in general and suggest neural structures for movement encoding in particular. | de |