CIF:Small:Toward a Modern Theory of Compression: Manifold Sources and Learned Compressors

过去我们通信基础设施的改进主要是通过提高我们在世界各地移动比特的能力来推动的。相比之下，人们很少关注如何使用数据压缩来减少必须首先进行通信的比特数。机器学习的最新进展表明，我们为维持电话会议或流媒体视频而必须发送的比特数比以前想象的要低得多。减少必须传输的比特数将使远程呈现更加逼真和身临其境。反过来，这可能导致旅行需求减少，包括通勤减少。由此带来的好处将是多方面的，包括减少碳排放，减少传染病的传播，以及为那些平衡工作与照顾孩子或老人的人提供更灵活的安排。该项目将有助于实现具有全方位远程呈现功能的透明、安全、廉价通信的巨大挑战。这个项目的推动力是基于人工神经网络的压缩器在多媒体源上获得的典型性能。虽然这些压缩机有吸引人的经验性能，但几乎没有理论来解释它们的性能或指出未来改进的方向。该项目将开发一个现代的数据压缩理论，这将有助于解释为什么这些压缩器的性能如此之好，估计它们的次优程度，并指出如何改进它们的方法。新的理论将集中在这些压缩器的两个最突出的方面，即它们对数据应用非线性变换的能力以及它们从源样本中学习的事实。该项目还将提高研究人员和信息理论成果的外部知名度，并通过女性和资深博士生的支持，实现更多元化的劳动力。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Past improvements in our communications infrastructure have been fueled primarily by increasing our ability to move bits around the world. In contrast, there has been less attention paid to how data compression can be used to reduce the number of bits that must be communicated to begin with. Recent advances in machine learning suggest that the number of bits that we must send to sustain a teleconference or stream a video is significantly lower than previously thought. Reducing the number of bits that must be transmitted would enable more realistic and immersive telepresence. This, in turn, may lead to a reduced need for travel, including reduced commuting. The resulting benefits would be manifold, including reduced carbon emissions, reduced spread of infectious disease, and more flexible arrangements for those balancing work with child- or elder-care. This project will contribute to the grand challenge of achieving transparent, secure, inexpensive communications with a full range of telepresence capabilities.The impetus for this project is the exemplary performance obtained by Artificial Neural Network-based compressors on multimedia sources. Although these compressors have attractive empirical performance, there is little theory to explain their performance or point the way toward future improvements. The project will develop a modern theory of data compression that will help explain why these compressors perform as well as they do, estimate their degree of suboptimality, and point the way toward how they can be improved. New theory will be developed focusing on the two most salient aspects of these compressors, namely their ability to apply nonlinear transforms to the data and the fact that they are learned from source samples. The project will also increase the external visibility of researchers and results in information theory and lead to more diverse workforce through the support of female and veteran Ph.D. students.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.