Organic optoelectronic neural networks

近年来，基于人工神经网络的机器智能取得了巨大的进步，几乎进入了科技、经济和日常生活的各个领域。然而，该领域目前的增长很大程度上依赖于不断增长的计算能力消耗，因此也就是电力消耗。这种对更大、更快系统的日益增长的需求是不可持续的，即使目前的重点是为MI过程开发定制的硬件。今天的数据中心已经消耗了全球总发电量的2%左右。这一数字正在呈指数级增长；IBM研究副总裁穆凯什·哈雷在2019年推断，到2040年，神经网络消耗的电力可能会超过世界发电量。因此，我们必须迫切地寻找从根本上新的计算原理来驱动MI。这个问题的一个有希望的解决方案是在人工神经网络中使用光而不是电子作为信息的主要载体。在光神经网络(ONN)中，光的波特性--相干和叠加--可以简化“矩阵乘法”运算(MI中计算量最大的运算)，从而提供了一种新的途径来极大地提高计算速度，同时极大地降低了功耗。当信息通过“前馈”神经网络的多层时，这种非线性函数被应用于每个神经单元，充当矩阵乘法各层之间的“垫圈”。原则上，任何非线性光学元件都可以起到AF的作用。然而，在实践中，由于损耗、缺乏灵活性和误差积累，使用纯光AFS实现大型ONN是具有挑战性的。在这里，我们将使用有机半导体器件来提供激活功能，通过光电二极管和OLED电路在光学层之间转换和传递信号。这将允许我们在每一层调节信号并纠正可能的错误，同时仍然利用光传播的优势来进行计算昂贵的步骤。智能手机中的OLED显示屏可以包含数百万个发射器，因此这一概念有可能扩展到能够执行非常复杂的计算任务的非常大的ONN。这个项目是两个小组的合作项目。圣安德鲁斯大学的PI是有机半导体光电子学领域的领导者，牛津PI拥有世界领先的光学计算硬件专业知识。圣安德鲁斯团队将开发“激活芯片”集成阵列，将自动对焦应用于多个光学单元。牛津集团将把这些激活芯片整合到适用于各种应用的ONN系统中。特别是，将开发一个概念上新颖的用于计算机视觉的ONN系统。该系统将允许神经网络直接“看到”和解释物体，而不需要将图像转换为电子形式。这样的系统将具有超低延迟，并可在自动驾驶汽车、遥感和智能机器人中找到应用。我们还将使用激活芯片来实施牛津集团的创新方法，直接训练ONN，这种方法不涉及数字模拟，因此既快又抗错误。

In recent years, machine intelligence (MI) based on artificial neural networks has made enormous progress, entering almost all spheres of technology, economy and our everyday life. However, much of the field's current growth is reliant on an ever-increasing consumption of computational power, and as a consequence electrical power. This growing demand for larger and faster systems is unsustainable, even with the current focus on developing bespoke hardware for MI processes. Today's data centres already consume about 2% of the total power generated worldwide. This number is growing exponentially; IBM vice president of research, Mukesh Khare, extrapolated in 2019 that the power consumed by neural networks could exceed the world's electricity production by 2040. We must therefore urgently look for fundamentally new computational principles to drive MI.A promising solution to this problem is to use light, rather than electrons, as the primary carrier of information in artificial neural networks. In optical neural networks (ONNs) the wave properties of light - coherence and superposition - can streamline the "matrix multiplication" operation (the most computationally expensive operation in MI), thereby offering a new route to greatly enhance computational speeds, with dramatically lower power consumption.This project aims to advance a crucial component of the ONN: the activation function (AF). This nonlinear function is applied to each neural unit as information passes through the multiple layers of a "feedforward" neural network, serving as a "gasket" between the layers of matrix multiplications. In principle, the AF role can be played by any nonlinear optical element. In practice, however, implementation of large ONNs with purely optical AFs is challenging due to losses, lack of flexibility and error accumulation. Here we will use organic semiconductor devices to provide the activation function, with circuits of photodiodes and OLEDs transforming and transferring the signal between optical layers. This will allow us to condition the signal at each layer and correct for possible errors, while still exploiting the advantages of light propagation for the computationally expensive steps. OLED displays in smart phones can contain millions of emitters, and so the concept is potentially scalable to very large ONNs capable of performing very complex computational tasks. The project is a collaboration of two groups. The PIs at the University of St Andrews are leaders in organic semiconductor optoelectronics and the Oxford PI possesses world-leading expertise in optical computing hardware. The St Andrews group will develop the "activation chips" - integrated arrays applying the AF to multiple optical units. The Oxford group will incorporate these activation chips into ONN systems suitable for various applications. In particular, a conceptually novel ONN system for computer vision will be developed. This system will allow a neural network to "see" and interpret objects directly, bypassing the need for converting an image into an electronic form. Such a system will have ultra-low latency and could find applications in autonomous vehicles, remote sensing and intelligent robotics. We will also use the activation chips to implement the Oxford group's innovative approach to direct training of ONNs, which does not involve digital simulation and hence is both faster and more robust to errors.