CSR: Small: Evolution of Computer Vision for Low Power Devices, Breaking its Power Wall and Computational Complexity

计算机视觉对物体识别和分类的准确性已经超过了人类的能力。脑启发卷积神经网络(CNN)模型的采用以及通过现代图形处理单元(GPU)训练和执行这些复杂网络的能力是这一进步的支柱。然而，在计算要求、内存使用和功耗方面，CNN解决方案要求极高。与此同时，计算机视觉的许多有趣的应用--如小型机器人、广泛的网络物理系统和物联网上的许多智能设备--都受到资源的限制。该项目旨在大幅降低基于CNN的VISION的计算复杂性、平均案例分类能力和延迟，使其能够部署到更广泛的平台。从社会的角度来看，这项研究通过吸收研究生、本科生、少数族裔和女性学生来加强乔治梅森大学(GMU)的研究、教育和多样性，并丰富了GMU提供的几门课程。本研究项目的目标如下：(1)将基于CNN的学习模型重新定义为迭代卷积神经网络(ICNN)学习模型，该模型允许早期分类并通过各种阈值机制允许早期终止，并开发一个框架来使用从早期迭代中提取的上下文知识来指导和减少未来迭代的计算。(2)开发了一种近似ICNN协处理器，通过探索ICNN创造的新的近似机会来支持内存和逻辑上的近似，并增强了ICNN在其预期功能之外的调整和学习近似硬件行为的能力。

The accuracy of computer vision for object recognition and classification has surpassed human capabilities. Adoption of brain-inspired Convolutional Neural Network (CNN) models and the ability to train and execute these complex networks by modern graphical processing units (GPUs) are the backbone of this progress. However, in terms of computational requirement, memory usage, and power consumption, the CNN solutions are extremely demanding. Meanwhile, many interesting applications of computer vision - such as small robotics, a wide range of Cyber-Physical Systems, and many smart devices on the Internet of Things - are resource constrained. This project aims to substantially lower the computational complexity, the average-case classification power and the latency of CNN-based vision, enabling its deployment to a much wider range of platforms. From a societal viewpoint, this study enhances the research, education, and diversity at George Mason University (GMU) by involving graduate, undergraduate, minority and female students, and enriches several courses that are offered at GMU.The goals of this research project are as follows: (1) Reformulating the CNN-based learning model into an Iterative Convolutional Neural Network (ICNN) learning model that allows early classification and permits early termination via various thresholding mechanisms and developing a framework to use the contextual knowledge that could be extracted from earlier iterations to guide and reduce the computation of future iterations. (2) Developing an approximate ICNN coprocessor that supports approximation in memory and logic by exploring new approximation opportunities created by ICNN, and enhancing the ICNN to adjust and learn the approximate hardware behavior in addition to its intended functionality.