Experimental Investigation of Time Series Classification using a Self-Pulsing Microring Resonator Network

Technical Summary: Experimental Investigation of Time Series Classification using a Self-Pulsing Microring Resonator Network

Problem Statement
Photonic neuromorphic computing offers significant advantages in power efficiency and parallel processing but often struggles to realize scalable nonlinearity and long-term memory without resorting to energy-intensive Optical-to-Electrical-to-Optical (OEO) conversions. While phase-change materials (PCMs) have been proposed for all-optical memory, they introduce high power losses and reset difficulties. Silicon microring resonators (MRRs) offer a potential solution by providing intrinsic nonlinearity and volatile memory through free-carrier and thermo-optic effects. However, there is a need for a robust, in-depth analysis of how coupled MRR networks perform as reservoirs for machine learning (ML) tasks, specifically regarding their ability to handle complex time-series data derived from images without extensive digital preprocessing.

Methodology
The authors experimentally investigated a photonic integrated circuit comprising a 64-node (8x8) network of coupled silicon microring resonators. The system operates as a Reservoir Computing (RC) platform where the physical hardware acts as a fixed, untrained recurrent neural network, and only a linear readout layer is trained.

• System Dynamics: The MRR network is driven into a nonlinear regime using input optical powers in the milliwatt range. This induces self-pulsing behavior (oscillations with periods ranging from ~1 to tens of microseconds) due to the interplay of free-carrier and thermal effects.
• Data Encoding: Images from the MNIST (handwritten digits) and Fashion-MNIST (clothing items) datasets (28×28 pixels) were flattened into 784-point time sequences. These sequences were encoded into optical intensity using an Arbitrary Waveform Generator (AWG) and an electro-optic modulator. Three encoding strategies were tested: "Normal" (background=0 power), "Inverse" (background=max power), and "Box" (background=40% max power).
• Dimensionality Expansion: To exploit the reservoir's capabilities, the system utilized multiple physical dimensions:
  • Spatial: Signals were acquired from six distinct output ports.
  • Spectral: Ten different laser frequencies (192.68–192.86 THz) were used.
  • Power: Multiple input power levels (1.1–5.5 mW) were varied.
• Classification Strategies:
  1. Full Time-Series: The entire temporal response of the network was recorded and fed into a linear classifier.
  2. Single-Pixel (Optical Memory): Classification was performed on a single time sample (pixel) of the output, leveraging the network's inherent memory to retain information about previous inputs without digital storage.
• Baselines: Performance was compared against a baseline linear classifier trained on raw digital data, an "out-of-resonance" photonic baseline (linear transformation only), and digital Multilayer Perceptrons (MLPs) with comparable parameter counts.

Key Contributions
This work expands upon previous studies of MRR networks by:

1. Distinct Mechanism: Utilizing complex dynamics from coupled self-pulsing in multiple MRRs for long-term memory, rather than relying on lossy PCMs.
2. Parameter Exploration: Systematically investigating the impact of output port count, encoding modality, and input sampling rates on ML accuracy.
3. Single-Pixel Inference: Demonstrating a classification setting where inference requires no digital memory of the signal history, relying solely on the optical memory of the MRR network.
4. Minimal Preprocessing: Using full datasets with minimal preprocessing, achieved via optical modulation rather than digital processing.
5. Computational Cost Analysis: Providing a direct comparison with digital ANNs regarding the number of trainable parameters and nonlinear operations required to achieve similar accuracies.

Results

• MNIST Classification: Using a linear readout on full time-series representations from multiple ports, the system achieved a peak test accuracy of 96.49% (using 5 ports, inverse encoding, 20 ns timestep). This significantly outperformed the baseline linear classifier on raw data (92.03%) and the out-of-resonance photonic baseline (92.12%).
• Fashion-MNIST Classification: The system achieved 85.81% accuracy, a modest improvement over the baseline (85.17%), attributed to the increased complexity of the dataset and noise from the photonic transformation.
• Comparison with Digital ANNs: To match the photonic system's 96.49% accuracy, a digital ANN required 120 hidden neurons (95,410 parameters) and 130 nonlinear operations. The photonic system achieved this with a linear readout requiring only 10 nonlinear operations (softmax) and approximately 61,260 trainable parameters (due to oversampling).
• Single-Pixel Classification: When classifying based on a single time sample (pixel), the photonic network achieved 68.57% accuracy for MNIST and 53.13% for Fashion-MNIST. This was a substantial improvement over the out-of-resonance baseline (~24-25%), demonstrating that the network's optical memory allows a single pixel to encode information about the entire image history.

Significance and Claims
The paper claims that the proposed MRR network provides a robust, compact (0.15 mm²), and energy-efficient (few mW input power) alternative to digital neural networks for time-series classification. The authors emphasize that the system's ability to perform dimensionality expansion through physical ports, wavelengths, and power levels allows a simple linear readout to achieve high accuracy.

Crucially, the work demonstrates that the inherent volatile memory and nonlinearity of coupled MRRs can replace complex digital memory and nonlinear activation functions in specific edge computing scenarios. The "single-pixel" result highlights the potential for ultra-low-latency inference where digital memory buffers are undesirable or unavailable. The authors position this not as a competitor to state-of-the-art digital CNNs for raw image classification, but as a proof-of-concept for efficient, hardware-native reservoir computing that reduces the computational overhead (parameters and nonlinear operations) required for specific tasks.