Collaborative Research: High Performance Cellular Simultaneous Recurrent Network based Pattern Recognition

这是一位研究人脸识别和图像识别问题的神经网络研究人员和一位研究基于忆阻器的新型电子芯片的研究人员共同提出的建议。近年来，卷积神经网络在图像识别方面创造了新的世界纪录，由其他大学通过NSF的EFRI/COPN主题资助。有时，这些系统在这些任务中的表现超过了人类。在神经网络方面，这个团队计划使用一类更通用的神经网络--细胞同时神经网络(CSRN)--来解决人脸识别中的基准挑战，在这个领域，计算机的表现还没有超过人类。CSRN可以被视为卷积网络的推广，以增加一种实时重现或反馈，一种已知对生物大脑的能力至关重要的重现。在电子硬件方面，这项提议解决了延续摩尔定律的一个关键挑战。计算芯片的速度预计不会像过去那样快速增长，但由于光刻技术的突破和最近在忆阻器方面的工作，我们仍然可以预期在一块芯片上实现数千甚至数百万个活跃处理器的进展。为了充分利用这一新兴的新能力，需要做出新的努力，将设备工作和系统级工作整合在一起，开发真正实用的新体系结构。如果成功，这个项目可能是这一努力向前迈出的重要一步。用于记忆的忆阻器已经得到了工业界的充足资金，但扩展到主动处理和学习更多的是一种高风险的突破性活动。该项目还包括教育和宣传的重要组成部分，包括开发激励K-8儿童的系统。

This is a collaborative proposal between a neural network researcher, addressing the issues of face recognition and image recognition in general, and a researcher on a new class of electronic chip based on memristors. In recent years, new world records have been set in image recognition by convolutional neural networks, funded at other universities through the EFRI/COPN topic at NSF. At times, those systems have outperformed humans in those tasks. On the neural network side, this team plans to use a more general class of neural networks, the Cellular Simultaneous Neural Network (CSRN), to address benchmark challenges in face recognition where computers have yet to outperform humans. The CSRN may be viewed asa generalization of the convolutional network to add a kind of real-time recurrence or feedback, a kind of recurrence which is known to be crucial to the powers of biological brains. On the electronic hardware side, this proposal addresses a crucial challenge in continuing Moore's Law. The speed of computing chips is not expected to grow as fast as it did in the past, but thanks to breakthroughs in lithography and the recent work in memristors, we can still expect progress towards thousand or even millions of active processors on a chip. In order to make full use of this emerging new capability, new efforts are needed to integrate device work and systems level work together, in developing new architectures of real practical use. If successful, this project could be an important step forward in that effort. Memristors for use in memory are already being well-funded by industry, but the extension to active processing and learning is more of a high riskbreakthrough activity. This project also includes a substantial component of education and outreach, including development of systems to stimulate K-8 children.