Yongkie Wiyogo, "Processing of sensory stimuli with recurrent neural networks", 2009.

Recurrent neural networks, one of many neural network concepts, are used in our work to process sensory stimuli received by an intelligent car. The hypothesis underlying this thesis is that the intrinsic properties of recurrent neural networks can be exploited in order to achieve a robust and error-free environment perception. By using dynamic neural fields, a subclass of recurrent neural networks, we can perform information processing of the autonomous car’s environment. The stability behavior of this mechanism is however still problematic if we want to apply it in technical systems. Thus, in this thesis, we employ the Lyapunov theorems for analyzing the dynamic behavior of dynamic neural fields. We show that stationary and non-stationary states of dynamic neural fields can be observed by visualizing the Lyapunov functional for dynamic neural fields. With this approach, we perform experiments to observe the behavior of dynamic neural fields for different types of stimuli. We also perform exploratory investigations to obtain early predictions of instabilities based on the analysis of the Lyapunov functional. By finding the inflection points in the Lyapunov functional, we find that we can indeed make sufficiently early predictions about instabilities.