Intelligent Holographic Nanostructures Can Reach 400 EXAFlops and Beyond Human Brain Neural Density

An AI optical circuit has been made with a neural density equal to 1/400 of human brain. The circuit is trained to process information through unpowered all-optical inference at the speed of light with a computational power more than ten orders of magnitude larger than electronic processors. In future, its neural density is expected to be 10 times that of human brain.

Machine learning is already used by millions every day to unlocking smartphones through facial recognition or passing through AI-enabled automated security checks at airports and train stations. Sensors collect optical information to feed it to a neural network in a computer. The new developments can make these sensors and AI microscopic.

Using a state-of-the-art laser 3D-nanoprinting technology, the researchers made optical perceptrons with a neuron density of over 500 million neurons per square centimeter. The nanoscale feature size of these smart optical elements pushes the upper limit for the computational power for the nanoprinted decryptors lies at 400 ExaFLOPS (10^18 FLOPS, floating operations per second). This is over 100,000 times more than integrated photonic hardware.

Above – a) Through computer machine learning training, the optical machine learning decryptor (MLD) acquires the capabilities of identifying a single decryption key (symmetric decryption, top) or entire classes of decryption keys (asymmetric decryption, bottom), and decoding a multitude of messages using a single decryptor element. b The decryption system can be considered a diffractive neural network for optical inference. Each layer of the network consists of N × N artificial neurons, secondary sources of waves (details in Supplementary Methods). c Schematic of an MLD integrated with a CMOS chip. The nanoscale MLD is physically 3D printed by GD-TPN (d), a nanofabrication method that gives precise control over the MLD neuron dimensions in the lateral and axial directions (e), achieving axial nanostepping of 10 nm

Printing the perceptrons directly on CMOS imaging chips they outperform current optical methods.

There are application in security check, medical diagnostics, automatic driving, satellite image processing, etc.

The machine learning decryptor MLD) is a single diffractive element capable of scattering and directionally focusing each of a multitude of images given as input and of mapping them into a specific output. Once printed, the MLD can optically perform the inference tasks of a single-layer perceptron, mapping a variety of images on a sensor, effectively realizing the functionalities of decryption.

The cointegration of our MLDs directly on CMOS chips allows further analysis of the output image collected at the detector plane in the electronic domain, which has been shown to be an energy-efficient method of hybrid optoelectronic image classification achieving accuracies up to 98.71%.

Computer-based machine learning training
The compact decryption system can be considered a diffractive neural network working in transmission mode. They modeled the MLD system on a computer to perform the training.

Light Science and Applications – Nanoprinted high-neuron-density optical linear perceptrons performing near-infrared inference on a CMOS chip.

Optical machine learning has emerged as an important research area that, by leveraging the advantages inherent to optical signals, such as parallelism and high speed, paves the way for a future where optical hardware can process data at the speed of light. In this work, we present such optical devices for data processing in the form of single-layer nanoscale holographic perceptrons trained to perform optical inference tasks. We experimentally show the functionality of these passive optical devices in the example of decryptors trained to perform optical inference of single or whole classes of keys through symmetric and asymmetric decryption. The decryptors, designed for operation in the near-infrared region, are nanoprinted on complementary metal-oxide–semiconductor chips by galvo-dithered two-photon nanolithography with axial nanostepping of 10 nm achieving a neuron density of over 500 million neurons per square centimeter. This power-efficient commixture of machine learning and on-chip integration may have a transformative impact on optical decryption, sensing, medical diagnostics and computing.

SOURCES- University of Shanghai for Science and Technology, Nature Light Science and Applications
Written By Brian Wang,

3 thoughts on “Intelligent Holographic Nanostructures Can Reach 400 EXAFlops and Beyond Human Brain Neural Density”

  1. "Each layer, is an N x N…." Seems it's multi-layer.
    "[plus] secondary sources of waves" This must be how they implement a nonlinear response.
    And yeah, the physical instantiation can't learn; it seems: "They modeled the MLD system on a computer to perform the training"

    Great if you ever have a NN system that never needs to be updated.
    Unlike any real complex system ever.
    Maybe it could have a use case for small modular image processing functions?
    A drop-in, "super resolution", photonic chip for your GFX card/camera/TV.
    Or edge detetection/motion extraction pre-processor for a more complex system.
    Small well defined functions that can be trained "perfectly" and then left alone.

  2. Well – looks like a hardcoded static model from reading this in a minute.
    Only for inference. No learning or training.
    Good performance but if it's static, a new device has to be printed or manufactured for every application or change.

  3. It's hard to guess what they're doing from the article. It sounds like a holographic system, which would prevent them from doing anything nonlinear. But they say it's implementing perceptrons, which inherently require a nonlinear operation.

    And they say it's doing "decryption" (as in cryptography?), but they also say it's only a single layer of perceptrons, which wouldn't be able to do any real cryptographic operations.

    And they say it's machine learning. But they also say it's a hologram nanoimprinted on a chip, which wouldn't be able to learn.


Comments are closed.