Study on Electoral College based Deep Learning and Its Applications

在理论方面，我们将研究基于选举人团的深度学习的稳定性。其最浅显的形式“选举人团”（Electoral College）在美国总统选举中已经使用多年，并应用于图像处理、模式识别、信息检索和其他政治选举等领域。我们的模型将是人工智能研究中第一个说明一类深度学习在地区/县大小、层数和每层相邻区域重叠率方面的稳定性的模型。我们的理论将指导应用程序如何选择适当的层数，适当的区域大小以及每层相邻区域的适当重叠率，以便使用基于选举团的深度学习来提高模式识别和信息检索方法的性能。

On the theoretical front, we will study the stability of the Electoral College-based Deep Learning. Its shallowest form, Electoral College, has been used for many years in US presidential elections, and applied in areas such as image processing, pattern recognition, information retrieval, and other political elections. Our model will be the first in Artificial Intelligence research that illustrates the stability of a class of Deep Learning with respect to region/prefecture size, number of layers, and overlapping rates of neighboring regions in each layer. Our theory will guide applications on how to choose an adequate number of layers, an adequate region size as well as adequate overlapping rates of neighboring regions for each layer, for using Electoral College-based Deep Learning to improve the performance of pattern recognition and information retrieval approaches. Developments made on the theoretical front will be applied in two important fields: (1) A face image set-based surveillance system will be developed. This system will combine proven active infrared illumination technology for pupil location with a mathematical approach developed in this research program to match unknown faces to known faces on watch lists. The system is expected to be a significant improvement, when compared to any appearance-based face recognition system, in recognition rate vs. false acceptance rate, and in being disguise-proof. (2) An intelligent textual information retrieval system will be developed. This system will employ a novel approach, where a document is treated as a physical object which can be described from different "view angles", for document representation so that each document can be divided into overlapped blocks. A vector space model will be employed as the basic strategy for matching a query with "block" in a document, and the Electoral College-based Deep Learning will be employed in making retrieval decisions. The system is expected to have much better performance in precision vs. recall compared to other approaches. These systems can be further developed for use in national security systems and internet intelligent document retrieval systems respectively.