Convolution is a fundamental process in deep learning (DL), especially in the field of computer vision (CV). It forms the basic building block of convolutional neural networks (CNNs), enabling models to automatically and efficiently learn hierarchical features from grid-like data such as images. However, this process consumes a significant amount of energy.
The developed chip integrates miniature Fresnel lenses (ultra-thin optical lenses etched directly onto silicon) with laser light to perform convolutions with much lower energy consumption and higher processing speeds. In prototype model tests, the chip achieved an accuracy of about 98% in classifying handwritten digits, compared to traditional electronic chips.
This research represents a critical step toward the widespread adoption of optical AI computing in everyday chips, as chip manufacturers are already using optical elements, which could facilitate the broad integration of this new technology in the near future. This optical chip signifies a promising leap towards faster and more environmentally friendly AI systems.
