Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit (2008.11659v1)

Abstract: Application-specific optical processors have been considered disruptive technologies for modern computing that can fundamentally accelerate the development of AI by offering substantially improved computing performance. Recent advancements in optical neural network architectures for neural information processing have been applied to perform various machine learning tasks. However, the existing architectures have limited complexity and performance; and each of them requires its own dedicated design that cannot be reconfigured to switch between different neural network models for different applications after deployment. Here, we propose an optoelectronic reconfigurable computing paradigm by constructing a diffractive processing unit (DPU) that can efficiently support different neural networks and achieve a high model complexity with millions of neurons. It allocates almost all of its computational operations optically and achieves extremely high speed of data modulation and large-scale network parameter updating by dynamically programming optical modulators and photodetectors. We demonstrated the reconfiguration of the DPU to implement various diffractive feedforward and recurrent neural networks and developed a novel adaptive training approach to circumvent the system imperfections. We applied the trained networks for high-speed classifying of handwritten digit images and human action videos over benchmark datasets, and the experimental results revealed a comparable classification accuracy to the electronic computing approaches. Furthermore, our prototype system built with off-the-shelf optoelectronic components surpasses the performance of state-of-the-art graphics processing units (GPUs) by several times on computing speed and more than an order of magnitude on system energy efficiency.

• The paper introduces a diffractive processing unit that achieves 270.5 TOPs/s and 5.9 TOPs/J, outperforming conventional GPUs.
• It leverages optical diffraction with electronic programmability to support diverse neural network architectures including feedforward and recurrent networks.
• The study demonstrates superior classification performance on benchmarks like MNIST, heralding future breakthroughs in optoelectronic neuromorphic computing.

Reconfigurable Diffractive Processing Unit for Large-Scale Neuromorphic Optoelectronic Computing

This paper introduces a novel optoelectronic computing paradigm utilizing a Reconfigurable Diffractive Processing Unit (DPU) that achieves high model complexity and versatility in supporting various neural networks. The DPU is designed to enhance capabilities in neuromorphic computing by exploiting the advantageous characteristics of optical processing to address the limitations inherent in electronic computing systems.

The authors propose that existing optical computing architectures, though promising, are limited by their inability to support diverse AI algorithms and their lower model complexity, which impacts their broader application viability. To this end, the paper outlines a solution in the form of a DPU, which combines optical and electronic advantages to form a versatile and programmable optoelectronic processor capable of high-speed data modulation and efficient network parameter updating.

The DPU leverages optical diffraction for a majority of its computational operations, integrated with electronically controlled programmability. This leads to significant improvements in computing speed and energy efficiency when compared with state-of-the-art electronic components such as GPUs. The authors showcase the superior performance of the DPU in implementing both feedforward and recurrent neural networks for tasks like high-speed handwritten digit and human action video classification, achieving superior inference capabilities and comparable accuracy levels to electronic computing.

Key Insights and Implications

The primary achievement of this research is the development of a DPU that supports millions of neurons, offering a high degree of flexibility in deploying different neural network architectures. The paper demonstrates that the DPU can be reconfigured to support a wide array of optical neural network configurations, enabling real-time dynamic adjustments to accommodate a range of applications including autonomous driving and edge computing.

Several compelling findings were reported:

• The DPU achieved a computing speed of 270.5 TOPs/s and system energy efficiency of 5.9 TOPs/J, outperforming contemporary GPUs by over 9 times in speed and more than an order of magnitude in energy efficiency.
• The classification accuracy of optoelectronic feedforward neural networks designed using the DPU surpassed the benchmark level achieved by LeNet-4 on the MNIST dataset and reached competitive levels on the Weizmann and KTH datasets.

These empirical results emphasize the potential of the DPU to stake a significant claim in high-performance neuromorphic computing applications, potentially eclipsing the limitations of traditional electronic processors as Moore's law reaches its physical limits.

Future Prospects

The research signifies a progressive step towards achieving more complex, efficient, and adaptable neuromorphic processors. The authors speculate that the DPU's performance could be further enhanced by integrating more advanced optoelectronic components, utilizing techniques such as spatial and spectral multiplexing, and embracing multi-wavelength operations. Long-term implications hinge on the intersection of optoelectronic innovations alongside improvements in training techniques, potentially influencing the development of robust AI accelerators.

This research catalyzes forward-thinking on the design of next-generation AI systems, underscoring the necessity to innovate beyond the conventional electronic paradigm to better exploit the potential of light-based computing. By integrating adaptive training methods, the researchers propose a path forward in overcoming experimental deviation issues, which traditionally hamper the efficacy of optical computational systems.

In conclusion, this paper offers a substantive contribution to the optoelectronic computing field, indicating significant room for growth and a potential shift in AI computational frameworks towards light-based solutions. This trajectory not only promises enhanced performance metrics but also serves to guide future exploration into the amalgamation of optics and electronics in large-scale neural computing architectures.