Inducing Features from Visual Noise using Statistical Machine Learning Techniques

对于想要了解视觉系统的心理学家和想要模拟其能力的计算机科学家来说，了解人类如何区分视觉对象是一个核心问题。以前的工作已经开发了用于确定图像位置和空间频率的技术，这些位置和频率对人类在分类和识别任务中的表现最为重要。然而，这些技术不能告诉我们重要的图像特征是否在单个步骤中被检测到，例如，通过将视觉对象与单个内部表示（例如模板）进行匹配，或者通过组合多个检测器的结果。在美国国家科学基金会的支持下，来自马萨诸塞州大学阿默斯特分校的安德鲁·科恩博士将开发一种新的模型，即多独立模板归纳模型（MITIM），旨在回答这些问题。在嘈杂图像中进行的单词识别等任务的人类行为实验表明，人们并不总是全面地识别物体。相反，它们似乎独立地检测片段，然后联合收割机将这些结果结合起来。如果这在人脸或物体分类和识别等领域是真的，那么理解这些能力需要了解独立检测到的部分及其相对重要性。拟议的MITIM算法将使用人工智能研究中的统计技术来处理人类图像分类数据，并发现用于对象识别的独立模板。获得的见解将有助于阐明人类的视觉性能，并帮助科学家和工程师更好地了解如何建立计算机系统，可以复制的性能，功能是越来越重要的网络搜索，机器人和安全应用程序。这一奖项是作为2006财政年度数学科学优先领域的数学社会和行为科学（MSBS）特别竞争的一部分。

Understanding how humans differentiate between visual objects is a central problem for psychologists who want to understand the visual system and for computer scientists who want to emulate its abilities. Previous work has developed techniques for determining the image locations and spatial frequencies most responsible for human performance in classification and identification tasks. These techniques, however, cannot tell us if the important image features are detected in a single step, for example, by matching the visual object to a single internal representation such as a template, or by combining the result of multiple detectors. With support from the National Science Foundation, Dr. Andrew Cohen from the University of Massachusetts, Amherst, will develop a new model, the Multiple Independent Template Induction Model (MITIM), designed to answer such questions.Human behavioral experiments on tasks such as word recognition in noisy images have suggested that people do not always recognize objects holistically. Rather, they appear detect pieces independently and then combine those results. If this is true in domains such as face or object classification and recognition, understanding these abilities requires knowledge of the independently detected parts and their relative importance. The proposed MITIM algorithm will use statistical techniques from artificial intelligence research to process human image classification data and discover the independent templates used in object recognition. The insights gained will help illuminate human visual performance and help scientists and engineers better understand how to build computer systems that can replicate that performance, functionality that is increasingly important in web search, robotic, and security applications.This award was supported as part of the fiscal year 2006 Mathematical Sciences priority area special competition on Mathematical Social and Behavioral Sciences (MSBS).