Theoretical simulations of dynamical systems for advanced reservoir computing applications

Abstract: There are computational problems that are simply too complex and cannot be handled by traditional CMOS technologies due to practical engineering limitations related to either fundamental physical behavior of devices at small scales, or various energy consumption issues. The field of unconventional computation has emerged as a response to these challenges. Up to date unconventional computation encompasses a plethora of computing frameworks, such as neuromorphic computing, molecular computing, reaction-diffusion computing, or quantum computing, and is ever increasing in its scope. This thesis is biased towards developing sensing applications in the unconventional computing context. This initiative is further extended towards developing novel machine learning applications. The possibility of building intelligent dynamical systems that collect information and analyze it in real-time has been investigated theoretically. The basic idea is to expose a dynamical system to the environment one wishes to analyze over time. The system operates as an environment sensitive reservoir computer. Since the state of the reservoir depends on the environment, the information about the environment one wishes to retrieve gets encoded in the state of the system. The key idea exploited in the thesis is that if the state of the reservoir is highly correlated with the state of the environment  then the information about the environment can be inferred with a modest engineering overhead. A typical dynamical system is assumed to be a network of environment sensitive elements. Each element can be something simple, but taken together, the elements acquire collective intelligence that can be harvested. These ideas have been examined theoretically (and verified experimentally) by simulating various networks of environment-sensitive elements: the memristor, the capacitor, the constant phase element and the organic field effect transistor element. The simulations were done in the context of ion sensing, which is an extremely complex, many-body, and multi-scale modeling problem.