eduzhai > Applied Sciences > Engineering >

Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: Application-specific optical processors have been considered disruptivetechnologies for modern computing that can fundamentally accelerate thedevelopment of artificial intelligence (AI) by offering substantially improvedcomputing performance. Recent advancements in optical neural networkarchitectures for neural information processing have been applied to performvarious machine learning tasks. However, the existing architectures havelimited complexity and performance; and each of them requires its own dedicateddesign that cannot be reconfigured to switch between different neural networkmodels for different applications after deployment. Here, we propose anoptoelectronic reconfigurable computing paradigm by constructing a diffractiveprocessing unit (DPU) that can efficiently support different neural networksand achieve a high model complexity with millions of neurons. It allocatesalmost all of its computational operations optically and achieves extremelyhigh speed of data modulation and large-scale network parameter updating bydynamically programming optical modulators and photodetectors. We demonstratedthe reconfiguration of the DPU to implement various diffractive feedforward andrecurrent neural networks and developed a novel adaptive training approach tocircumvent the system imperfections. We applied the trained networks forhigh-speed classifying of handwritten digit images and human action videos overbenchmark datasets, and the experimental results revealed a comparableclassification accuracy to the electronic computing approaches. Furthermore,our prototype system built with off-the-shelf optoelectronic componentssurpasses the performance of state-of-the-art graphics processing units (GPUs)by several times on computing speed and more than an order of magnitude onsystem energy efficiency.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...