Quantum Transport Reservoir Computing

TL;DR

The paper introduces a practical route to quantum reservoir computing by implementing it with quantum transport in mesoscopic conductors, exploiting universal conductance fluctuations as the nonlinear reservoir. Gate voltages modulate UCF, and the resulting current serves as the observable output, circumventing the need for multiple copies or strong readout back-action. Using a scattering-matrix framework and Landauer conductance, the authors demonstrate two benchmarks—spoken-digit recognition and NARMA2 time-series forecasting—showing high accuracy and tunable performance. The approach emphasizes compatibility with existing fabrication, direct electrical readout, and resilience to measurement back-action, enabling on-chip, nanoscale quantum RC with potential impact on mesoscopic quantum information applications.

Abstract

Reservoir computing (RC), a neural network designed for temporal data, enables efficient computation with low-cost training and direct physical implementation. Recently, quantum RC has opened new possibilities for conventional RC and introduced novel ideas to tackle open problems in quantum physics and advance quantum technologies. Despite its promise, it faces challenges, including physical realization, output readout, and measurement-induced back-action. Here, we propose to implement quantum RC through quantum transport in mesoscopic electronic systems. Our approach possesses several advantages: compatibility with existing device fabrication techniques, ease of output measurement, and robustness against measurement back-action. Leveraging universal conductance fluctuations, we numerically demonstrate two benchmark tasks, spoken-digit recognition and time-series forecasting, to validate our proposal. This work establishes a novel pathway for implementing on-chip quantum RC via quantum transport and expands the mesoscopic physics applications.

Figures (8)

• Figure 1: Architecture of a new quantum RC type. The input $u(t)$ is initially encoded in the reservoir. By adjusting parameters $\{\theta_i\}$ of the quantum system action $S$ using $u(t)$, the quantum system produces the internal state $w(t)$ nonlinearly. This process maps $u(t)$ nonlinearly into a high-dimensional space. The weight matrix $W_{\text{out}}$ is subsequently optimized linearly to produce the final $y(t)$.
• Figure 2: (a) Schematic of a quantum reservoir using UCF with two gates. Parameters $\{\theta_i\}$ here are various gate voltages applied on the sample, and the internal state $w(t)$ is formed by the resulting current values. (b) UCF induced by varying gate voltages. Encodings '0' and '1' are represented by gate voltages $0$ and $V$, respectively. Various encodings result in distinct conductance values across a wide range. Performance of our reservoir can be precisely tuned via gate voltages. The relevant parameters are: $M=30$, $l_{el}=100/3$, $l=l_{el}$, 10 impurities.
• Figure 3: (a) Audio waveform of digit 0 from the NIST TI46 database. (b) Cochleagram with 64 frequency channels and 40 time steps, converted from (a) using Lyon’s passive ear model and setting a threshold. (c) Training and testing accuracy as functions of iterations when $V=5.9$. The highest training and testing results are 99.6$\%$ and 94$\%$, respectively. (d) Confusion matrix visualizes the distribution of predicted digits compared to the actual spoken digits when $V=5.9$. (e) Training and testing accuracy as functions of $V$.
• Figure 4: (a) Training and testing inputs and outputs when the training process utilizes 40 distinct values of $V$. Only 100 frames are shown in these figures for clarity. (b) NRMSE trends as functions of size.
• Figure 5: For the second sample with four gates: (a) Training and testing accuracy as functions of iterations when $V=0.1$. The highest training and testing results are 99.3$\%$ and 96$\%$, respectively. (b) Confusion matrix visualizes the distribution of predicted digits compared to the actual spoken digits when $V=0.1$. (c) Training and testing accuracy as functions of $V$. For the third sample with five gates: (d) Training and testing accuracy as functions of $V$.