Biologically Inspired Intelligent Classifiers

这个项目将推进智能代理设计方面的知识。它有两个目标。第一个目标是在自动模式分类器的构建中，为自动化特征发现和进化学习（两种受生物学启发的长期学习形式）开发数学见解和过程。第二个目标是在大规模图像数据库和时变视觉信号的自动分析仪中结合自动特征发现和进化。该项目将通过建立在首席研究员和他的学生的自适应超平面放置、窗口训练（随机近似的扩展形式）、松弛分支和界搜索以及遗传特征选择的技术基础上，努力实现这些目标。该项目将在手写数字字符的三阶段神经分类器、大型数据库中医学图像的树分类器以及数字化视频中运动物体的自适应分割中实现和测试长期学习的结果机制。该项目将有助于智能机器的长期学习技术和多传感器机器人视觉，通过自动发现来自不同传感器的高度区分的特征组合。

This project will advance knowledge about design of intelligent agents. It has two objectives. The first objective is the development of mathematical insights and procedures for automating feature discovery and learning by evolution -- two biologically inspired forms of long-term learning -- in the construction of automatic pattern classifiers. The second objective is the incorporation of automated feature discovery and evolution in automatic analyzers of large-scale image data banks and time-varying visual signals. This project will work to achieve these objectives by building on the Principal Investigator's and his students' techniques of adaptive hyperplane placement, window training (an extended form of stochastic approximation), relaxed branch-and-bound search, and genetic feature selection. The project will implement and test the resulting mechanisms of long-term learning in three-stage neural classifiers of handwritten numerical characters, tree classifiers of medical images in a large data bank, and adaptive segmenter of moving objects in digitized video. This project will contribute to the technology of long term learning in intelligent machines and to multi-sensor robotic vision by automating the discovery of highly discriminating combinations of features from diverse sensors.