Abstract
Reservoir computing has garnered significant attention for its efficiency in processing temporal signals, while the proposed next-generation reservoir computing (NG-RC) further enhances computational efficiency. The analog in-memory computing architecture fundamentally reduces the data transfer operations between data processing units and storage units. This study proposes a novel NG-RC implementation scheme based on memristor crossbar arrays. Specifically, the nonlinear feature vector operation is mathematically converted into matrix multiplication, enabling in situ computation within memristor crossbar arrays. The unique monomial of the vector outer production can be extracted by reading the current of the specific column line in array, facilitating efficient NG-RC implementation. Under ideal memristor device conditions, the proposed scheme demonstrates excellent performance in the prediction and inference tasks, validating its feasibility. Additionally, simulation experiments considering non-ideal memristor characteristics reveal that the programming error is the critical factor affecting the system performance. However, by selecting the appropriate memristor device, the proposed NG-RC implementation scheme shows comparable performance to digital implementations, confirming the effectiveness of the proposed scheme.