CAREER: Characterizing feature selectivity and invariance in deep neural architectures

本CAREER提案的目标是帮助阐明使鲁棒目标识别成为可能的原则。物体识别是所有生物必须解决的问题，从单细胞生物到人类。尽管基于化学事件、光或声波的识别物理信号不同，但分析这些事件的计算要求似乎是相似的。具体来说，有两个主要属性，任何系统调解鲁棒对象识别必须具有。第一个性质被称为“不变性”。它使神经元对从不同角度观察到的同一物体产生相似的反应。第二个特性被称为“选择性”。选择性要求神经元对可能非常相似的物体（如不同的面孔）产生不同的反应，即使是从相似的角度呈现。让检测器保持不变但不选择性或选择性但不不变是很简单的。难点在于制造既具有选择性又具有不变性的检测器。这个职业项目将开发统计方法，同时表征神经元的不变性和它们对环境中特定特征的选择性。所开发的方法将具有三个显著特征。首先，它将有可能恢复新的不变性类型，而无需事先假设任何给定神经元或大脑区域的主要不变性类型是什么。其次，它们将使描述不完全和近似类型的不变性成为可能。第三，这些方法将面向自然感官环境的典型刺激，这些刺激具有丰富的对象，并从感觉处理的各个阶段引起神经元的强烈反应。发展方法的这三个特性将使同时研究不同区域内和跨区域的多个神经元成为可能，而不需要调整对特定神经元或大脑区域的刺激。将所开发的方法应用于大脑中介导视觉和听觉物体识别的神经元的反应，将有助于揭示大脑中感觉处理的共同原理，并可能最终导致人工识别系统的改进设计，包括感觉假体。这项研究将纳入教育和推广活动涉及K-12学生，本科生和研究生。教育部分将有助于整合计算机科学、物理学和神经科学方面的知识，培养精通这些学科的新一代科学家。与当地学校和博物馆的联系，以及在线课程的创建，将有助于接触到当地和世界各地的各种各样的学生。

The goals of this CAREER proposal are to help elucidate the principles that make robust object recognition possible. Object recognition is a problem that must be solved by all living organisms, from single-cell organisms to humans. Although the physical signals for recognition based on chemical events, light or sound waves are different, the computational requirements for analyzing these events appear to be similar. Specifically, there are two main properties that any system that mediates robust object recognition must have. The first property is known as "invariance." It endows neurons with a similar response to the same object observed from different viewpoints. The second property is known as "selectivity." Selectivity requires that neurons produce different responses to potentially quite similar objects (such as different faces) even when presented from similar viewpoints. It is straightforward to make detectors that are invariant but not selective or selective but not invariant. The difficulty lies in making detectors that are both selective and invariant. This CAREER project will develop statistical methods for simultaneously characterizing both the invariance properties of neurons and their selectivity to specific features in the environment. The developed methods will have three distinguishing characteristics. First, it will be possible to recover new types of invariance without any prior assumptions of what the dominant type of invariance is for any given neuron or brain region. Second, they will make it possible to characterize imperfect and approximate types of invariance. Third, the methods will be geared towards stimuli typical of the natural sensory environment that are rich in objects and elicit robust responses from neurons from all stages of sensory processing. These three properties of the developed methods will make it possible to simultaneously study multiple neurons both within and across different regions, without the need to adjust stimuli to a particular neuron or brain region. Application of the developed methods to responses of neurons that mediate visual and auditory object recognition in the brain will help reveal the common principles of sensory processing in the brain and may ultimately lead to improved designs of artificial recognition systems, including sensory prostheses.This research will be integrated into education and outreach activities involving K-12 students, undergraduate and graduate students. The educational component will help integrate knowledge acquired in computer science, physics, and neuroscience, training a new generation of scientists that are proficient in these disciplines. Outreach to local schools and museums, as well as the creation of an online course will help reach a diverse range of students both locally and worldwide.