SGER: Popularizing Neural Processes: A Project to Place anOptoelectronic Neural System in every Wallet

9617121贾维迪人工神经网络模拟生物神经系统，其特征是被称为神经元的大规模相互连接的处理元件。这一过程是通过训练和调整神经元之间相互连接的强度来执行的。神经网络系统是解决难以用数学描述的问题的理想选择。尽管神经网络在解决复杂的计算问题方面前景光明，但它们在日常应用中还没有得到广泛的应用，它们的应用主要局限于军事项目。神经网络的广泛商业化将极大地惠及该领域，并将为研发提供更多资金。开发一种实用、低成本的光电神经系统，利用神经网络的计算潜力，将促进神经网络和光电技术的进一步商业化。最近的研究发现了用于人脸识别和其他生物识别应用的新硬件和软件，这些硬件和软件既独特地适合开发神经处理的特征，又具有广泛商业化的潜力。这项提议旨在开发一种紧凑、低成本的光电处理器，以实现用于人脸识别的神经系统。人脸识别和分类是一项艰巨的任务，因为人脸的外观由于头部视角、面部表情、光照、发型等的不同而不断变化。人脸识别问题需要存储大量的面部特征数据库，以成功地实现人脸图像分类所需神经元之间的大规模互联。提出的系统可以使用光记录材料，如光聚合物，来存储身份证上个人的大型面部数据库，如驾照。为了进行检查，持卡人的实时图像被显示在光电设备上，并与存储在卡上的光致聚合物膜中的面部特征进行比较。这种比较是由光电神经芯片进行的，它执行准确分类所需的计算。为了增加安全性，面部特征可以通过相位编码进行光学加密，以防止未经授权的人复制身份证5。该系统的优点之一是使用光学材料，可以在相对较小的区域内存储大量的面部信息，很容易安装在卡上。此外，面部信息由可以对大量数据进行并行计算的光束并行读出，因此可以在短时间内进行识别。将开发神经算法，以提供可靠的人脸识别方法。挑战在于生产非常低的出错概率算法、显示信息的低成本输入输出设备以及紧凑的光学系统架构和设计。我们提出将基于非线性滤波的光电神经网络与有监督的感知学习算法相结合用于实时人脸识别。第一层用非线性联合变换相关器(JTC)2-4光学实现，第二层由于隐含层神经元的数目较少而用电学实现。该系统利用输入的面部图像序列进行训练，并且能够实时地对输入的人脸进行分类。基于非线性联合变换的特点，该系统具有以下特点：易于光学实现和系统对准；系统可以集成到低成本的紧凑型原型中；通过更新输入的参考图像(权重)来训练系统，该参考图像(权重)可以存储在电子或光学存储器中，不需要产生滤波器或全息图；采用傅立叶平面的非线性变换，系统对输入图像的不良变化具有鲁棒性，具有良好的分辨灵敏度和对噪声的鲁棒性；系统具有平移不变性。该处理器可以使用商业上可获得的光电子设备，并且可以被构建为低成本的紧凑型系统。将进行计算机模拟和光学实验结果，以确定系统在识别输入面部图像时的错误概率。通过对研究中的输入图像进行时分复用，我们希望在使用多个输入图像的情况下，分类错误的概率可以降低到零。***

9617121 Javidi Artificial neural networks mimic biological neural systems and are characterized by massively interconnected processing elements called neurons. The processing is performed by training and adjusting the strength of the interconnection between neurons. Neural network systems are ideal for solving problems that are difficult to describe mathematically. Even though neural networks hold a great promise for solving complex computational problems, they have not been utilized extensively in day-to-day applications, and their application has been limited mainly to military projects. Widespread commercialization of neural network would greatly benefit the field and would make available more funds for R&D. Development of a practical, low-cost optoelectronic neural system that would take advantage of the computational potential of neural networks would facilitate further commercialization of both neural network and optoelectronic technology. Recent research has discovered new hardware and software for an application for face recognition and other biometrics that is both uniquely suited to exploit the characteristics of neural processing and has the potential for widespread commercialization. This proposal aims to develop a compact, low-cost optoelectronic processor to implement a neural system for face recognition. Face recognition and classification is a difficult task because the facial appearance is constantly changing due to different head perspectives, facial expressions, different illuminations, hair styles, etc. The face recognition problem requires storing a large data base of facial features to successfully implement the massive interconnection between the neurons required to classify facial images. The proposed system can use optical recording materials such as photopolymers to store the large facial data base for an individual on an ID card, such as a driver's license. For inspection, a live image of the person carrying the card is displayed on an optoelectronic device and is compared to the facial features stored in the photopolymer film on the card. This comparison is performed by an optoelectronic neural chip which carries out the computation necessary for an accurate classification. For additional security, the facial features can be optically encrypted by phase encoding to prevent reproduction of the ID cards by unauthorized people 5 . One of the advantages of the proposed system is that using optical materials, large amounts of facial information can be stored on a relatively small area, fitting easily on the card. Also, the facial information is read out in parallel by light beams which can perform parallel computation on large arrays of data, and therefore the recognition can be performed in a short period of time. Neural algorithms will be developed to provide reliable face recognition methods. Challenges are in the area of producing very low probability of error algorithms, low-cost input-output devices to display the information, and compact optical systems architectures and designs. We propose to use a nonlinear filter based optoelectronics neural networks associated with a supervised perception learning algorithm for real-time face recognition. The first layer is implemented optically using a nonlinear joint transform correlator (JTC) 2-4 and the second layer is implemented electronically because of the small number of the hidden layer neurons. The system is trained with a sequence of input facial images and is able to classify an input face in real-time. Based on the characteristics of the nonlinear JTC, the proposed system has the following features: it is easy to implement optically and is robust in terms of system alignment; the system can be integrated into a low-cost compact prototype; the system is trained by updating the reference images (weights)in the input which can be stored in electronic or optical memories and no filters or holograms need to be produced; using nonlinear transformation in the Fourier plane, the system is robust to ill umination variations of the input image, has a good discrimination sensitivity and is robust to noise; and the system is shift invariant. The processor may use commercially available opto-electronics devices and can be built as a low cost compact system. Computer simulations and optical experimental results will be performed to determine the probability of error of the system in identifying input facial images. By using time multiplexing of the input image under investigation, we hope that using more than one input image, the probability of error for classification can be reduced to zero. ***