Single-chip photonic deep neural network for instantaneous image classification (2106.11747v1)

Abstract: Deep neural networks with applications from computer vision and image processing to medical diagnosis are commonly implemented using clock-based processors, where computation speed is limited by the clock frequency and the memory access time. Advances in photonic integrated circuits have enabled research in photonic computation, where, despite excellent features such as fast linear computation, no integrated photonic deep network has been demonstrated to date due to the lack of scalable nonlinear functionality and the loss of photonic devices, making scalability to a large number of layers challenging. Here we report the first integrated end-to-end photonic deep neural network (PDNN) that performs instantaneous image classification through direct processing of optical waves. Images are formed on the input pixels and optical waves are coupled into nanophotonic waveguides and processed as the light propagates through layers of neurons on-chip. Each neuron generates an optical output from input optical signals, where linear computation is performed optically and the nonlinear activation function is realised opto-electronically. The output of a laser coupled into the chip is uniformly distributed among all neurons within the network providing the same per-neuron supply light. Thus, all neurons have the same optical output range enabling scalability to deep networks with large number of layers. The PDNN chip is used for 2- and 4-class classification of handwritten letters achieving accuracies of higher than 93.7% and 90.3%, respectively, with a computation time less than one clock cycle of state-of-the-art digital computation platforms. Direct clock-less processing of optical data eliminates photo-detection, A/D conversion, and the requirement for a large memory module, enabling significantly faster and more energy-efficient neural networks for the next generations of deep learning systems.

• The paper demonstrates a single-chip photonic deep neural network that processes images in less than one clock cycle while achieving over 93.7% accuracy in classification tasks.
• It employs an optical-electronic hybrid architecture with grating coupler arrays, nanophotonic waveguides, and integrated photodiodes to perform both linear and nonlinear computations.
• By eliminating the need for analog-to-digital conversion, the research paves the way for real-time, energy-efficient AI systems with scalable photonic neural networks.

Integrated Photonic Deep Neural Networks for Image Classification

The paper presents a groundbreaking advancement in the domain of deep neural networks (DNNs) by introducing an integrated end-to-end photonic deep neural network (PDNN) that offers instantaneous image classification capabilities. This innovation leverages the properties of photonic integrated circuits to circumvent the limitations of traditional clock-based digital processors, providing significant improvements in computation speed and energy efficiency.

System Architecture and Implementation

The proposed PDNN is designed to process optical waves directly, bypassing the need for conversion to the electrical domain. Images are projected onto a grating coupler array, with the subsequent optical waves being channeled into nanophotonic waveguides. As light propagates through the chip, it interacts with neuron layers, each of which performs both linear and nonlinear computations essential for classification tasks.

The neurons perform linear computations in the optical domain, while nonlinear activation, specifically the ReLU function, is realized through opto-electronic mechanisms. This combined optical and electronic process is achieved using components such as integrated photodiodes and micro-ring modulators, contributing to the PDNN's ability to handle more sophisticated network layers and deep networks.

Performance and Results

The PDNN chip achieved above 93.7% accuracy for 2-class and 90.3% accuracy for 4-class handwritten letter classification tasks. Notably, the computation time was drastically reduced to less than a single clock cycle, compared to state-of-the-art digital systems. The computation paradigm of "computation-by-propagation" implemented in the PDNN eliminates traditional bottlenecks such as photo-detection and analog-to-digital conversion, contributing to its efficiency and speed.

Implications and Future Directions

This research indicates a significant shift in deep learning implementation, suggesting that photonic systems could outpace traditional digital counterparts, especially for tasks requiring high-speed, energy-efficient processing. Practical applications of this technology include real-time pattern recognition and event detection systems, which could benefit from the rapid processing capabilities inherent in photonic systems.

The potential scalability of the PDNN to more complex networks and higher resolution images is another notable aspect. Strategies for managing routing complexity and facilitating effective scaling—such as utilizing multiple photonic routing layers or tiling techniques—are discussed, emphasizing the adaptability and future potential of photonic neural networks.

Conclusion

The development of a photonic deep neural network capable of real-time image classification signifies a notable advancement in optical computing. This approach challenges the dominance of electronic processing units in deep learning, providing a pathway towards novel, efficient, and scalable AI systems. Future research could focus on further improving the scalability, diversification, and integration of photonic neural networks, potentially leading to their wider adoption across various domains of artificial intelligence and beyond.