The paper presents a novel approach to reservoir computing, a brain-inspired machine learning framework designed for processing temporal data by mapping inputs into high-dimensional spaces. Traditional physical reservoir computers (PRCs) often rely on homogeneous device arrays, which use input encoding methods and large stochastic device-to-device variations to achieve nonlinearity and high-dimensional mapping. These methods, however, come with high pre-processing costs and limitations in real-time deployment.
This study introduces a heterogeneous memcapacitor-based PRC that leverages internal voltage offsets to enable both monotonic and non-monotonic input-state correlations. These correlations are crucial for efficient high-dimensional transformations. The proposed approach eliminates the need for input encoding methods, thereby reducing pre-processing costs and enhancing real-time applicability.
The efficacy of this novel PRC is demonstrated through two key experiments. First, the model predicts a second-order nonlinear dynamical system with an extremely low prediction error of 0.00018. Second, it predicts a chaotic Hénon map, achieving a low normalized root mean square error (NRMSE) of 0.080. These results underscore the power of distinct input-state correlations facilitated by the internal voltage offsets in the memcapacitors.
Furthermore, the approach is generalized to other neuromorphic devices that lack inherent voltage offsets by applying external offsets to realize various input-state correlations. This generalization broadens the applicability of the proposed method across different types of neuromorphic devices.
The paper concludes that the heterogeneous memcapacitor-based PRC represents a significant advancement in the field of reservoir computing. By achieving unprecedented performance without the need for input encoding methods, this approach marks a major milestone towards the development of high-performance, full in-materia PRCs. This innovation holds promise for more efficient and effective processing of temporal data in various applications.
