CRCNS: Representation and Computation in Natural Vision

在自然环境中，物体在各种光照条件、姿势、背景和与其他物体并置的情况下被观察。人工视觉系统的不变性足以适应这种变化，但它的选择性却永远不够。真实图像的丰富结构提供了大量的偶然安排，其中许多导致系统错误地检测到不存在的对象。另一方面，高度选择性的系统同时在面对自然变化时也很容易错过检测。相比之下，人类和动物的视觉系统能够在广泛的观察条件下准确地看到东西——生物系统是如何既具有选择性又具有不变性的呢？对这个问题的追求导致了一个关于复杂细胞和其他在腹侧视觉通路中普遍存在的不变细胞类型的类似问题。他们的优势似乎是他们的弱点：视觉系统如何可能在不变性的基础上建立选择性？复杂细胞的模型给出了一种解释。复杂细胞和其他不变细胞类型，由于其非线性响应特性，必然具有功能连接，因此这些细胞与它们输入的一个通常很小的子集具有功能连接。这种承诺是间接的，因为它取决于接受野中的特定模式。功能连通性是迄今为止提出的复杂细胞接受场特性的几乎所有非线性模型的一个可论证的数学性质。更重要的是，这些观察得出的结论是，具有重叠接受野的成对细胞将显示出功能性的共同输入。这也是间接的，事实上，当各个接受域的模式“契合在一起”——对应于一个更大的整体的各个部分时，功能性共同输入就会很高。这些观察结果为不变性与选择性之间的困境提供了一个解决方案：适当地组合在一起的片段产生了高度的功能性共同输入，这表现为在其他不变性表示之间的统计依赖，最有可能以部分同步的形式出现，从而向视觉通路中更深层次的细胞发出部分组成的信号。为了对选择性/不变性困境的答案进行实验验证，研究人员采用新的统计和方法技术来研究不变细胞的接受场特性的新问题，并测量具有重叠接受场的不变细胞的联合统计中的新变量。

Representation and Computation in Natural VisionIn natural environments, objects are viewed under a wide variety of lighting conditions, poses, backgrounds, and juxtapositions with other objects. Artificial vision systems that are sufficiently invariant to accommodate such variations are never sufficiently selective. The rich structure of real images offers a multitude of chance arrangements, many of which cause systems to falsely detect an object that is not there. On the other hand, systems that are highly selective are at the same time highly prone to missed detections in the face of natural variability. The visual systems of humans and animals, in contrast, are able to see accurately under a wide range of viewing conditions--how is it that biological systems are both selective and invariant?The pursuit of this question leads to an analogous question about complex cells and other invariant cell types that are ubiquitous in the ventral visual pathway. Their strength would appear to be their weakness: How is it possible for the visual system to build selectivity out of invariance? Models of complex cells suggest an explanation. Complex and other invariant cell types, by virtue of their nonlinear response characteristics, necessarily possess a functional connectivity whereby these cells become functionally connected to a generally small subset of their inputs. This commitment is circumstantial, inasmuch as it depends on the particular pattern in the receptive field. Functional connectivity is a demonstrable mathematical property of virtually all of the non-linear models put forward to date for complex-cell receptive-field properties. What is more, these observations lead to the conclusion that pairs of such cells that possess overlapping receptive fields will demonstrate a functional common input. This too is circumstantial, and in fact functional common input is high exactly when the patterns in the respective receptive fields "fit together"---correspond to pieces of a larger whole. These observations suggest a solution to the dilemma of invariance versus selectivity: pieces that fit properly together generate a high degree of functional common input, which manifests itself by a statistical dependence between otherwise invariant representations, most likely in the form of partial synchrony, thereby signaling a composition of parts to cells deeper in the visual pathway. In search of experimental confirmation of this proposed answer to the selectivity/invariance dilemma, the investigators employ new statistical and methodological techniques to study new questions about the receptive-field properties of invariant cells, and to measure new variables in the joint statistics of invariant cells with overlapping receptive fields.