Massively Parallel Amplitude-only Optical Processing System for Machine Learning

Machine learning has become a driving force in modern technology. However, existing technology cannot keep up with its demand, due to limitations by capacitive charging of wires, system architecture that stores and handles data, and many more. Deep learning algorithms like, Convolution Neural Networks (CNN), extracts specific features of interest, using linear mathematical operations (convolutions). These convolution layers consume majority (~80%) of the compute resources during the process of making decisions based on observations or data.

Researchers at GW have developed a novel Amplitude-only Fourier-optical processor system, which is capable of processing large-scale (~1,000 x 1,000) dimension matrices in a single step with 100 microsecond- short latency. It is an optoelectronic system where low-power laser light is partnered by electronically configured Digital Micro mirror Devices (DMDs) in both object and Fourier plane. For DMDs, by individually controlling the 2 million programmable micro mirrors, it is possible to achieve reprogrammable operations for (near) real-time, which is about 100x lower latency compared to current GPU accelerators.

 

Figure 1. Shows 4-f optical processing systems used as convolutional layer of a neural network.

 

Advantages:

  • Low latency
  • Efficient power consumption
  • Neural network can attain a classification accuracy of 98% for MNIST (Modified National Institute of Standards and Technology database)

 

Application:

  • Image classification
  • Super-resolution imaging on unmanned aerial vehicle
  • High bandwidth communication in data centers
  • Accelerator for performing artificial intelligence tasks
Patent Information: