Selected from TechCrunch
Author: Devin Coldewey
Translation by Machine Heart
The speed of signal transmission between neurons in the brain is about 100 meters per second, while the speed of light is 300,000 kilometers per second. What if neuronal signals could also propagate at the speed of light? Researchers from the University of California, Los Angeles (UCLA) have used 3D printing technology to create solid-state neural networks and executed computations through hierarchical propagation of light diffraction, achieving handwritten digit image recognition. The relevant results have been published in the journal Science.
This idea seems novel, but it is quite natural. The operations performed in neural networks are linear, which corresponds well with the linear interactions of light diffraction. The concepts of neuron weights and activation values can also correspond to the amplitude and phase of light (which are adjustable). Additionally, solid-state optical diffraction computing has advantages such as low energy consumption, no heat generation, and execution at light speed (although the electric field propagation in traditional computer circuits is also at light speed, it does not directly correspond to the computational processes of neural networks). This research direction is still in its infancy, but if its advantages can be fully utilized, it may have a broad application prospect.
Today, machine learning is ubiquitous, but most machine learning systems are invisible: they optimize audio or recognize faces in images within a “black box.” Recently, UCLA researchers developed a 3D printed AI analysis system. This system is not only visible but also tangible. Unlike previous systems that analyzed data through digital adjustments, this system analyzes artificial intelligence through the diffraction of light. This novel and unique research result indicates that these “AI” systems can appear very simple.
We typically consider machine learning systems as a form of artificial intelligence, where the core involves a series of operations performed on a set of data, with each operation based on the previous one or fed into a loop. The operations themselves are not overly complex — although they are not simple enough to be calculated with pen and paper. Ultimately, these simple mathematical operations yield a probability that indicates how the input data matches the various patterns that the system has “learned” to recognize.
Usually, the computations required for each parameter update or inference in machine learning systems need to be performed on a CPU or GPU. Due to the current deep learning requirements for extensive parallel computations, GPUs have become a more widespread choice. However, even the most advanced GPUs are made of silicon and copper, and information must propagate in pulses along complex circuits. This means that whether executing new computations or repetitive computations, traditional GPUs will generate energy consumption.
Therefore, when these “layers” in deep learning have completed training and all parameter values are determined, they will still repeatedly compute and consume energy. This means that the 3D printed AI analysis system can be optimized after training its “layers,” without occupying too much space or CPU power. Researchers from UCLA indicate that it can indeed be solidified, with these layers themselves made from transparent materials, 3D printed layers imprinted with complex diffraction patterns that can manipulate light.
If this description makes you feel a bit dizzy, think of a mechanical calculator. Nowadays, digital calculations are completed in a computer’s logic in digital form. However, in the past, calculators needed to move actual mechanical parts to perform calculations — adding up to 10 would cause a change in the positions of the parts. To some extent, this “diffractive deep neural network” is similar: it uses and manipulates the physical representation of numbers rather than electronic representations. This means that if the model’s prediction process is solidified into a physical representation, it can significantly reduce energy consumption during the actual prediction process.
As the researchers state:
Each point on a given layer transmits or reflects the incident wave, which corresponds to an artificial neuron connected to other neurons in the next layer through optical diffraction. By altering the phase and amplitude, each “neuron” is adjustable.
“Our all-optical deep learning framework can execute various complex tasks at light speed, which can also be achieved by computer-based neural networks,” the researchers wrote in their paper describing their system.
To prove this, they trained a deep learning model to recognize handwritten digits. Once completed, they transformed the matrix math layers into a series of optical transformations. For example, one layer might increase a value by refocusing the light from both to a single area of the next layer — the actual computation is much more complex than this overview suggests.
By arranging millions of micro-conversions on the printed plate, light is input from one end and output from another structure, allowing the system to determine with over 90% accuracy whether it is a 1, 2, or 3.
Readers may wonder what the use of this is, as even the simplest three-layer perceptron can easily achieve over 95% accuracy in recognizing handwritten digits, while convolutional networks can achieve over 99% accuracy. This form currently has no practical use, but neural networks are highly flexible tools, and the system could potentially recognize letters instead of being limited to digits. This could enable optical character recognition systems to operate in hardware, requiring virtually no energy consumption or computation.
The real limitation lies in the manufacturing process: creating a high-precision diffraction plate capable of on-demand processing tasks is very challenging. After all, if precision is needed to seven decimal places, but the printed plate can only achieve three decimal places, that would be quite troublesome.
This is merely a proof of concept — there is no pressing demand for large digital recognition machines — but the idea is quite intriguing. This concept may impact camera and machine learning technologies — constructing light and data in the physical world rather than the virtual world. It may seem like a regression, but perhaps it’s just the pendulum swinging back.
Paper: All-optical machine learning using diffractive deep neural networks
Paper Link: http://science.sciencemag.org/content/early/2018/07/25/science.aat8084
Abstract: Deep learning has enhanced our ability to perform advanced reasoning tasks using computers. In this paper, we introduce a physical mechanism for executing machine learning, an all-optical diffractive deep neural network (D^2NN) architecture that can implement multiple functions according to passively diffractive layers designed based on deep learning. We constructed a 3D printed D^2NN to achieve image classification of handwritten digits and fashion products, as well as imaging lenses functioning in the terahertz spectrum. Our all-optical deep learning framework can compute various complex functions that traditional computer-based neural networks can also achieve at light speed, and it will develop new applications in all-optical image analysis, feature detection, and object classification. Furthermore, it allows for the design of new cameras and optical devices to utilize D^2NN for unique tasks.
Figure 1: Architecture of the diffractive deep neural network (D^2NN).
Figure 2: Testing experiment of the 3D printed diffractive deep neural network.
Figure 3: Diffractive deep neural network achieving handwritten digit recognition.
Original link: https://techcrunch.com/2018/07/26/this-3d-printed-ai-construct-analyzes-by-bending-light/
This article is translated by Machine Heart, please contact this public account for authorization.
✄————————————————
Join Machine Heart (Full-time Reporter / Intern): [email protected]
Submit or seek coverage: content@jiqizhixin.com
Advertising & Business Cooperation: [email protected]