Produced by Big Data Digest
Authors:Ning Jing, Wei Zimin
Have you ever thought about moving neural networks from computers into a piece of glass?
Using neural networks for image recognition and intelligent recommendations has become very common. In recent years, the increase in computing power and parallel processing has made it a very practical technology. However, at its core, it remains a digital computer, no different from other computer programs. Moreover, its demand for electrical energy is also increasing.
Taking handwritten digit recognition as an example, this is a classic introductory topic in deep learning: you need to go through the forward propagation of layers of neural networks and adjust the weights and biases of each layer of neurons using the BP algorithm supported by the principle of gradient descent to achieve the final recognition effect, the complexity of the calculations and parameter exchanges in between is considerable.
For a long time, researchers have been exploring faster and more energy-efficient methods to perform the complex calculations of neural networks.
Recently, researchers from the University of Wisconsin, MIT, and Columbia University released a new type of neural network that uses a special glass panel, simulating units in the network with linear and nonlinear materials. After training, it can perform the same tasks as ordinary neural networks.
The most interesting part is that this special glass panel requires almost no electricity; it only needs light to operate. Currently, the research team has used it to recognize grayscale handwritten digits with an accuracy rate of 79%.
This special glass contains precisely controlled inclusions, such as pores or impurities like graphene or other materials. When light shines on the glass, complex ripple patterns occur, and in one of the ten regions on the glass, the light becomes more intense. Each region corresponds to a digit.
For example, here are two examples of recognizing the handwritten digit “2” on the glass:
The research team stated that they designed very strict rules to create the relevant experimental equipment and will further improve the materials to enhance recognition efficiency under the premise of being as “scalable” as possible. The team also plans to create the network in 3D.
Here is the link to the related paper 👇
https://www.osapublishing.org/prj/fulltext.cfm?uri=prj-7-8-823&id=415059
Let’s take a detailed look with the digest bug at how this magical glass neural network is realized.
From ANN to Optical Neural Computing
Artificial Neural Networks (ANN) have been widely used in deep learning, but they require increasingly enhanced computing power from computers, prompting efforts to find faster and more energy-efficient alternative computing methods, a typical method being optical neural computing. This simulation computing method has minimal energy consumption, and more importantly, its inherent parallelism can greatly accelerate computing speed.
Most optical neural computing follows the architecture of digital neural networks, using layered feedforward networks, as shown in the figure below (a), where free-space diffraction or integrated waveguides are used as connections between activated neurons, similar to digital signals in ANN, and optical signals pass through the optical network once in the forward direction.
So what acts as the feedback for the BP algorithm?
It is the reflection of light that provides the feedback mechanism, resulting in rich wave physics. Here, optical reflection indicates the possibility of transcending the paradigm of layered feedforward networks to achieve artificial neural computation in a continuous and unlayered manner.
The figure below (b) shows the proposed Nanophotonic Neural Medium (NNM). The optical signal enters from the left side, and the output is the energy distribution on the right side of the medium. The computation is performed by the main material such as silica, with many inclusions; the inclusions can be pores or any other material with a different refractive index from the main medium, which strongly scatters light in both forward and backward directions.
Why have inclusions?What role do they play in optical neural computing?
The position and shape of inclusions correspond to the weight parameters in digital neural networks, and their sizes are typically sub-wavelength. Nonlinear operations can be implemented through inclusions made of dye semiconductors or graphene saturable absorbers, where they perform distributed nonlinear activation. These nonlinear designs consider rectified linear units (ReLU), which allow signals with intensity above a threshold to pass through and block signals with intensity below that threshold.
To better illustrate this behavior, figure (d) shows the output intensity of light with a wavelength of λ, thickness of λ/2, through the design of this nonlinear material, achieving a nonlinear function of the incident wave intensity, using this material as a nonlinear activation, as shown in light blue.
The following figure shows the operation of the NNM, where the two-dimensional (2D) medium is trained to recognize grayscale handwritten digits. The dataset contains 5000 different images, with representative images as shown in figure (a), where each image represented by 20×20 pixels is converted into a vector and then encoded as the spatial intensity of the input light incident from the left side.
Inside the NNM, nanostructures generate strong interference, and depending on the digit represented by the image, the light is guided to one of the ten output positions, where the position with the highest energy light intensity corresponds to the category determined as the final result. Figure (b) shows the process created by two different handwritten digits “2”; due to the different shapes of the handwritten digits, the field patterns produced by these two images are completely different. You can see that the shape of the yellow light in figure (b) is slightly different, but both show high light intensity in the same area at the output, correctly completing the recognition of the handwritten digit.
Similarly, figure (c) shows the recognition of two different shaped handwritten digits “8”, where strong light intensity is produced in another position among the ten regions. Here, the finite difference frequency domain (FDFD) method is used to solve the nonlinear wave equation to simulate the field. The size of the NNM is from 80λ to 20λ, where λ is the wavelength of light used to carry and process information.
For a test set consisting of 1000 images, the average recognition accuracy reached over 79%, with the reported accuracy limited due to strict constraints set by researchers during the optimization process of the manufacturing issues. These constraints keep the medium dense; otherwise, it would consist of sparse air and silica. By relaxing these requirements or using larger medium sizes, the final test accuracy can be further improved.
How Far from Laboratory to Industry?
This glass-based neural network seems to be still a prototype in the laboratory, but during the research process, the digest bug was pleasantly surprised to find that as early as 2015, a French startup was already implementing neural networks using this optical method.
This company, called LightOn, has an interesting slogan: “Changing the future of computing with light,” hoping to enhance computing speed, size, and power efficiency by simulating computing devices with optical instruments, and has launched a hardware coprocessor called the Optical Processing Unit (OPU). It aims to enhance some of the most computationally intensive tasks in machine learning. The OPU can be inserted into standard servers or workstations and accessed through a simple toolbox that can be seamlessly integrated into familiar programming environments, with comprehensive OPU prototypes provided to selected users via LightOn Cloud.
In the use case section, the digest bug also found that some of the practical applications of this company have gone beyond recognizing handwritten digits, including image and video recognition, recommendation systems, NLP, and other related systems.
LightOn official website:
https://www.lighton.ai/our-technology/
Although the relevant data is still some distance from current industry usage, for the computing field that pursues computing power and energy efficiency, it may indeed bring something different to this area.
Related reports:
https://hackaday.com/2019/07/16/neural-network-in-glass-requires-no-power-recognizes-numbers/#more-367116
Recommended Course | Data Science Training Camp Session 8
4 months, hands-on guidance through the entire process of data collection, cleaning, storage, analysis, visualization, machine learning modeling, big data!
Intern/Full-time Editor Reporter Recruitment
Join us and experience every detail of writing for a professional technology media outlet, growing alongside a group of the best people in the most promising industry. Located in Beijing, Tsinghua East Gate, reply to the dialogue page on the Big Data Digest homepage with “Recruitment” to learn more. Please send your resume directly to [email protected]
