BioCOMP: Biologically Inspired Computational Model for Perception

摘要：加州大学河滨分校获得一项资助，用于开发生物启发的视觉感知计算模型。感知系统接受原始的感觉输入，如数字照片/电影，并识别它们包含的对象。几十年来，人们一直在尝试开发感知系统，但收效甚微。这项研究的关键创新在于，它将神经科学家已经确定的指导自然知觉系统发展的生物学约束纳入到计算知觉系统的自动化发展中。该项目涉及一个跨学科的团队，以及一名计算机科学家和一名认知心理学家之间的密切合作。在这项研究中开发的系统学习以一种类似于生物系统的方式工作。该项目开发了一种新的范例，将特定领域的知识，在这种情况下，生物限制到进化计算中，以开发创新的视觉系统。该方法系统地解决了现实环境中目标检测/识别问题的复杂性和严重性。这项研究产生了计算创新，以允许开发能够利用发育神经心理学限制的进化学习系统。它们包括(A)协作式协同进化，它允许任务的组件在协作改进适应度的环境中进化；(B)智能交叉和变异算子，它在计算进化的几代中保留有效组件；以及(C)最小描述长度约束，它基于描述效率和拟合优度来选择算子。这些创新的目标，无论是集体的还是单独的，都是为了减少进化学习过程必须遍历的搜索空间的体积，以使其能够解决感知问题。该项目将使用几个公开可用的数据库来演示结果。

Abstract: The University of California Riverside is awarded a grant to develop biologically inspired computational models for visual perception. Perceptual systems take raw sensory input like digital photographs/movies and identify the objects they contain. Decades have been spent trying to develop perceptual systems, with only modest success. The key innovation of this research is that it incorporates the biological constraints neuroscientists have identified that guide the developmental of natural perceptual systems into the automated development of computational perceptual systems. The project involves an interdisciplinary team and a close collaboration between a computer scientist and a cognitive psychologist. The systems developed in this research learn to work in a way that is similar to biological systems. The project develops a new paradigm for incorporating domain-specific knowledge in this case, biological constraints into evolutionary computation to develop innovative visual systems. This approach systematically addresses the complexity and magnitude of the object detection/recognition problem in real-world environments. The research generates computational innovations to permit the development of evolutionary learning systems that can utilize developmental neuropsychological constraints. These include (a) cooperative coevolution that allows components of a task to evolve in an environment in which cooperation improves fitness, (b) smart crossover and mutation operators that retain effective components over generations of computational evolution, and (c) a minimum description length constraint that selects operators based on efficiency of description in addition to goodness of fit. The goal of these innovations, collectively and individually, is to reduce the volume of the search space that the evolutionary learning process must traverse to allow it to solve the perceptual problem. The project will use several publicly available databases to demonstrate the results.