Preview

PDF, English - main document Download (21MB) | Terms of use

Abstract

Inspired by the remarkable properties of the human brain, the fields of machine learning, computational neuroscience and neuromorphic engineering have achieved significant synergistic progress in the last decade. Powerful neural network models rooted in machine learning have been proposed as models for neuroscience and for applications in neuromorphic engineering. However, the aspect of robustness is often neglected in these models. Both biological and engineered substrates show diverse imperfections that deteriorate the performance of computation models or even prohibit their implementation. This thesis describes three projects aiming at implementing robust learning with local plasticity rules in neural networks. First, we demonstrate the advantages of neuromorphic computations in a pilot study on a prototype chip. Thereby, we quantify the speed and energy consumption of the system compared to a software simulation and show how on-chip learning contributes to the robustness of learning. Second, we present an implementation of spike-based Bayesian inference on accelerated neuromorphic hardware. The model copes, via learning, with the disruptive effects of the imperfect substrate and benefits from the acceleration. Finally, we present a robust model of deep reinforcement learning using local learning rules. It shows how backpropagation combined with neuromodulation could be implemented in a biologically plausible framework. The results contribute to the pursuit of robust and powerful learning networks for biological and neuromorphic substrates.