Design principles for optoelectronic light-scattering reservoir computing at the edge of chaos

Technical Summary: Design Principles for Optoelectronic Light-Scattering Reservoir Computing at the Edge of Chaos

Problem Statement
Physical reservoir computing (RC) offers an energy-efficient pathway for sequential cognitive inference by leveraging the intrinsic nonlinear dynamics of hardware substrates. While free-space light-scattering systems are attractive due to their parallelism and reconfigurability, a critical gap remains: practical design principles linking specific hardware control variables to computational performance have not been established. Unlike conventional recurrent neural networks (RNNs) where weights are learned via backpropagation, physical RC systems rely on fixed, physically determined dynamics. Consequently, performance is highly sensitive to the choice of substrate parameters, including the dynamical regime, input-reservoir coupling, and interconnectivity. While theoretical and simulation studies suggest an optimal operating regime exists at the "edge of chaos" (the boundary between stability and chaos), experimental confirmation in physical reservoirs has been limited and substrate-specific. This work addresses the lack of a systematic, quantitative design framework for reconfigurable optoelectronic light-scattering reservoirs.

Methodology
The authors implemented a free-space optoelectronic reservoir computing system using a 4-f imaging architecture. The system consists of a coherent laser source, a phase-modulating spatial light modulator (SLM) for input encoding and state projection, a second reconfigurable SLM acting as a scatterer to mix the optical field, and a CMOS camera for intensity detection. The reservoir state update is governed by the recurrence of the optical field intensity, mathematically modeled as:
$$\mathbf{h}(t) = \min(|\mathbf{W}_{rec} \exp(i\pi((\mathbf{I}-\tilde{\mathbf{W}}_{in})\mathbf{h}(t-1) + \tilde{\mathbf{W}}_{in}\tilde{\mathbf{x}}(t)))|^2, 1)$$
where $\mathbf{h}$ is the reservoir state, $\mathbf{x}$ is the input, $\mathbf{W}_{rec}$ represents the interconnectivity, and $\mathbf{W}_{in}$ represents the input coupling. Only the readout matrix $\mathbf{W}_{out}$ is numerically trained using Moore–Penrose pseudoinverse.

The study systematically mapped three physical control axes:

1. Reservoir Dynamics: Controlled by the camera exposure time ($T$), which modulates the norm of $\mathbf{W}_{rec}$ and the overall light intensity.
2. Input–Reservoir Coupling: Controlled by the density ($\rho$, the fraction of nodes coupled to input) and magnitude ($w$, the weight of nonzero entries) of the input matrix $\mathbf{W}_{in}$.
3. Reservoir Interconnectivity: Controlled by the diffusing angle ($\theta$) of the scatterer's phase profile, which determines the spread of the optical field across nodes.

The authors measured the maximal Lyapunov exponent ($\lambda_{max}$) to characterize dynamical stability and calculated the Memory Capacity (MC) to evaluate computational performance. These metrics were compared against numerical simulations to validate the physical model.

Key Contributions and Results
The paper establishes three quantitative design principles by identifying optima along the three control axes:

1. Edge of Chaos as a Universal Operating Point:
   By sweeping the camera exposure time, the authors observed a transition from stable ($\lambda_{max} < 0$) to chaotic ($\lambda_{max} > 0$) regimes. A distinct peak in Memory Capacity was found to coincide precisely with $\lambda_{max} \approx 0$. This experimental confirmation extends the "edge of chaos" hypothesis, previously verified in ion-gating and spin-wave reservoirs, to the photonic substrate. The study demonstrates that the MC peak is a property of the underlying dynamics rather than an artifact of noise or hardware imperfection.

2. Input–Reservoir Coupling Trade-off:
   With dynamics fixed at the critical point ($\lambda_{max} \approx 0$), the authors swept the input coupling parameters ($\rho$ and $w$). The results revealed a "density–magnitude trade-off": low coupling fails to register external information, while excessive coupling silences the reservoir's internal history. An optimal operating region was identified at $\rho \approx 21\%$ and $w \approx 0.25$, where input driving and recurrent dynamics are balanced.

3. Intermediate Interconnectivity:
   The study examined the effect of the scatterer's diffusing angle ($\theta$) on interconnectivity. Memory capacity exhibited a non-monotonic dependence on $\theta$, with a moderate maximum at $\theta = 0.2^\circ$ for a 10,000-node reservoir. This suggests that information propagation between nodes requires an intermediate breadth—neither too localized nor too diffuse.

Validation on Cognitive Tasks
Operating at the identified three-axis optimum ($\lambda_{max} \approx 0$, $\rho=21\%$, $w=0.25$, $\theta=0.2^\circ$), the reservoir was tested on two tasks:

• Chaotic Time-Series Prediction: The system successfully predicted the Mackey–Glass chaotic time series in "free-running" mode (where predictions are fed back as inputs), maintaining a bounded phase-space trajectory that closely resembled the ground truth. This performance was only achievable at the critical dynamics point; stable and chaotic regimes failed to reproduce the attractor.
• Spoken Digit Recognition: The reservoir achieved an 84.5% blind classification accuracy on the 10-class Speech Commands dataset (spoken digits). The confusion matrix showed strong diagonal dominance, with errors concentrated on acoustically similar digit pairs.

Significance and Claims
The authors claim that this work provides the first systematic mapping of the design landscape for reconfigurable optoelectronic light-scattering reservoirs. The primary significance lies in establishing a transferable framework rooted in substrate-independent concepts of criticality and balanced coupling, despite being stated in substrate-specific units.

The paper positions photonics as the third experimental substrate (after ion-gating and spin-wave systems) to confirm the edge-of-chaos hypothesis in physical reservoir computing. The authors emphasize that while their system is optoelectronic (involving an electronic round-trip for detection and re-encoding), the design principles regarding dynamics, coupling, and connectivity are expected to transfer to fully-optical implementations, such as cavity-based or integrated photonic platforms, where the physical control handles differ (e.g., cavity gain or intra-cavity modulators) but the underlying optimization landscape remains similar. The work does not claim to maximize bandwidth or minimize latency compared to fully-optical systems but rather to define the quantitative rules governing computational performance in light-scattering reservoirs.