Abstract
This work provides a comprehensive investigation into methods of informing neural networks with prior knowledge in the specific area of low latency signal emulation for virtual analog modelling, particularly with nonlinear circuits. It focuses on two novel neural methods, namely, the Forward and Inverse method. In the modelling of nonlinear circuits, as adherence to the rules of realtime safety and computational cost is often required, there is usually a compromise in the modelling of the nonlinear functions associated with the behaviour of the circuit. The canonical approach has generally been to use underlying assumptions about the behaviour to provide approximations such that realtime computation can be successfully achieved. On the other hand, with advent of blackbox methods such as neural networks, the problem lies in the cost of deploying a model with an increased parameter count, leading to problems where the model might be too large to fit on cache, slowing down inferencing times. Compounded onto this is the problem that in order for user parameters to be available, the model would either need to have an increased parameter count, or multiple copies of the model need to be deployed, further increasing the problems associated with blackbox modelling techniques. The primary objective of this research is to investigate the proposed informed neural methodologies, and their efficacy both in terms of accuracy and computational cost, while still maintaining the exposure of user parameter controls. This relies on the advantages of both whitebox and blackbox modelling techniques, and tries to see if there is a goldilocks zone for a “best of all worlds scenario”. The beginning of the work focuses on providing building blocks for the two novel neural methods to be built upon, and the latter half of the work focuses on the implementation of a possible application of the two neural methods and their evaluation. Lastly, two optimization techniques for the implementation and the inference of the neural methods are proposed. These four contributions complete the full picture surrounding the solutions associated with the virtual analog modelling using informed neural methods. Beyond this, the formulation of these methods provides a basis for further informed neural methods to be built in an attempt to provide the flexibility and approximation ability of neural networks, without the drawback of the lack of transparency of the blackbox approaches, with computational costs equivalent to, or sometimes better than, traditional whitebox approaches.