CAREER: Intersubband neurons for ultrafast optical neural networks

人工神经网络是一种模仿人脑操作的机器学习技术。这些网络通常是使用电子设备实现的，它们促成了许多最近的技术进步。使用光的神经网络-光学神经网络-可能执行更快的计算，可能以光速进行。然而，与电子学竞争的全规模光网络还没有得到证明，因为它们缺乏允许人工神经网络做出决策的关键元素。在这个项目中，将开发新的光学设备来填补这一缺失的拼图。通过在不同材料上生长原子薄层，第一个子带间神经元将被创造出来。这些设备将能够根据照射到它们身上的光量做出决定，它们最终将允许开发出超高速光学神经网络。这可以直接使许多领域受益，因为它可以直接加速许多计算任务。此外，它还可以实现完全不使用电子设备的信息处理！该计划将研究和教育融为一体，对社区产生了更广泛的影响。它将为南本德的一所中学开发一个光学外展项目，向学生介绍重要的概念，并让他们有机会看到真正的研究实验室的运作。它还将为未被充分代表的群体的本科生开发一个暑期研究项目，以及一个新的非线性光学研究生课程。技术描述：该项目的主要目标是开发用于信息处理的新的子带间光子器件，最终形成第一个能够高速运行的光神经网络。基于神经网络的深度学习使计算发生了革命性的变化。通过将线性矩阵乘法与非线性激活函数级联，深度神经网络可以学习许多任务。原则上，光学神经网络可以以光速进行计算，比电子网络快数千倍。不幸的是，虽然光在计算网络的线性部分方面很出色，但它却不能很容易地计算出非线性部分。光学非线性很快，但出了名的小。在这个项目中，将设计一种纳米结构，将光学元件与非线性电子元件混合在一起。该计划将利用子带间转换的物理原理来制造子带间神经元，这种非线性设备的运行速度预计会比现有设备快得多，而且光功率更低。几种新的设计策略已经被开发出来，它们可以实现低阈值、低功耗、高速的人工神经元，在这个项目中，他们将被实验演示和表征。还将使用一种新兴的材料系统在较短的波长下开发神经元，以提高这一概念的可扩展性和长期生存能力。这个项目的智力优势在于，它将为一种全新的光学神经网络方法奠定基础，这种方法无缝地融合了电子学的最佳特性和光子学的最佳特性。尽管子带间物理以前已经被用来制造信号源、探测器和传感器，但它们还没有对计算产生影响--这个程序将做到这一点。它将在光学、电子工程和计算机科学的交叉领域做出重要贡献。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

An artificial neural network is a machine learning technique that mimics the operation of a human brain. These networks are typically implemented using electronics, and they have been responsible for many recent technological advancements. Neural networks that use light—optical neural networks—could potentially perform calculations even faster, potentially at the speed of light. However, full-scale optical networks competitive with electronics have not been demonstrated, as they lack the critical element that allows artificial neural networks to make decisions. In this program, new optical devices will be developed that fill in this missing puzzle piece. By growing atomically-thin layers of different materials on one another, the first intersubband neurons will be created. These are devices that will be able to make decisions based on the amount of light that hits them, and they will eventually allow for ultrafast optical neural networks to be developed. This could directly benefit many fields, as it could provide direct speed-up of many computing tasks. In addition, it could allow for information processing that does not use electronics at all! This program integrates research and education, having broader impacts on the community. It will develop an optics outreach program for a middle school in South Bend, one that introduces students to important concepts and will allow them the opportunity to see a real research lab in action. It will also develop a summer research program for undergraduates from underrepresented groups, as well as a new graduate course on nonlinear optics.Technical description:The main goal of this program is to develop new intersubband photonic devices for information processing, ultimately culminating in the first optical neural networks capable of high-speed operation. Deep learning based on neural networks has revolutionized computation. By cascading linear matrix multiplications with nonlinear activation functions, a deep neural network can learn many tasks. In principle, optical neural networks could perform calculations at the speed of light, thousands of times faster than electronic networks. Unfortunately, while light is excellent at computing the linear part of the network, it cannot so easily compute the nonlinear part. Optical nonlinearities are fast but notoriously small. In this program, a nanostructure will instead be designed that blends an optical element with a nonlinear electronic element. This program will utilize the physics of intersubband transitions to make intersubband neurons, nonlinear devices expected to operate at speeds much faster than existing devices and with lower optical powers. Several novel design strategies have been developed that can implement low-threshold, low power consumption, high-speed artificial neurons, and in this program, they will be experimentally demonstrated and characterized. Neurons will be also be developed at shorter wavelengths using an emerging material system in order to improve the scalability and long-term viability of this concept. The intellectual merit of this program is that it will lay the groundwork for a completely new approach to optical neural networks, one that seamlessly blends the best features of electronics and the best features of photonics. Though intersubband physics have previously been exploited to make sources, detectors, and sensors, they have yet to make an impact in computation—this program will do just that. It will make important contributions at the intersection of optics, electrical engineering, and computer science.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.