The digit classification of silicon microring resonators shows the performance of reservoir computing with thermal nonlinearity

Recognizing handwritten numbers is a key issue in artificial intelligence, and researchers are continually investigating new approaches to improving accuracy and efficiency. Alongside Andrey A. Nikitin, Aa Ershov, Av Kondrashov, Alexander S. Smirnov, Sergey S. Kosolobov, and Anastasiya K. Zemtsova, we demonstrated an important step by successfully implementing DIGIT classification using photonic leisure computers based on thyricon sales computers. This innovative system utilizes the inherent nonlinearity of light within the resonator to create complex, higher-dimensional spaces that allow pattern recognition, effectively mimicking the behavior of neural networks. The team's work represents a promising pathway to compact, energy efficient, chip-scale artificial intelligence, offering potential alternatives to traditional electronic systems for pattern recognition and machine learning tasks.

This system offers a promising alternative to traditional electronic computing, compact and energy efficient computing systems for pattern recognition tasks. The team successfully classifies handwritten numbers using a specifically designed photonic reservoir, demonstrating the technology potential of real applications.

The team presents the first experimental investigation of a reservoir computer based on a single silicon microring resonator operating in the digit recognition task. The input layer consists of a laser and an electro-optical modulator, which encodes the light intensity applied to the resonator. This input signal is converted into a high-dimensional virtual space via thermal nonlinearity within the resonator. The resonator response records a record containing the photodetector and the oscilloscope. The goal is to demonstrate a compact, energy-efficient hardware implementation of reservoir computing for potential applications in machine learning and signal processing. This study exploits the reservoir computing paradigm to simplify recurrent neural network training by modifying recurrent connections and training only the output layer to reduce computational complexity. This study demonstrates the ability of photonic reservoirs to perform tasks such as nonlinear automatic regression and time series prediction, demonstrating the potential of this approach to building compact and efficient machine learning hardware. This paper highlights the advantages of photonic implementation over traditional electronic and other physical reservoir computing approaches. The team accomplishes this by exploiting the nonlinear response of the resonator to light. In particular, it achieves changes in transmission coefficients induced by frequency shifts and fluctuations in input power. This effect creates fading memory within the system. This is essential for processing information in a reservoir computing framework. Performance is evaluated through short-term memory and parity check testing to demonstrate the feasibility of this approach and achieve comparable capabilities to those obtained with other reservoir architectures.

This study includes detailed characterization of the linear and nonlinear transmission characteristics of the resonator, revealing a clear relationship between input capabilities and frequency shifts. For example, an increase in the input power of 2 dBM induces a 2.4GHz frequency shift, highlighting the intensity of the nonlinear effect. This work paves the way for hybrid manufacturing of photonic integrating circuits to combine lasers, resonators and photosectors to create a compact and efficient neurogenetic computing platform.