Robin Msiska
Reservoir computing is a computational framework in which a fixed non-linear dynamical system called a reservoir is used to transform intricate non-linear input data such they can be linearly separated at the output nodes using simple algorithms like linear regression. In this work, We implement a high-performance skyrmion mixture reservoir with multi-dimensional inputs. We use this system as a model to develop and study general performance metrics for the key reservoir computing characteristics of memory and non-linearity. These metrics, as opposed to standard reservoir computing metrics, are spatially resolved and can be computed concurrently from a single input signal which accelerates parameter searches to design effective reservoirs. Furthermore, we use the developed metrics to optimise our skyrmion reservoir and evaluate its performance using a spoken digit benchmark. The results show an overall model accuracy of 97.4% and a word error rate of <1%, among the best ever reported for in-materio physical reservoir computing