Although perceptual grouping has been widely studied, its mechanisms remain poorly understood. We propose a neural model of grouping that, through top-down control of its circuits, implements a grouping strategy involving both a connection strategy (which elements to connect) and a selection strategy (that defines spatiotemporal properties of a selection signal to segment target elements and facilitate identification). We apply the model to a letter discrimination task that investigated relationships among uniform connectedness and the grouping principles of proximity and shape similarity. Participants reported whether small circles formed a global letter E or H, and these circles could be connected by a line or be embedded in a matrix of squares. In the model, a good grouping strategy for this task consists of a connection strategy that connects circles but not squares for all conditions and a selection strategy that uses a selection signal of varying size, depending on whether squares were present. Consistent with empirical results, which were verified in two replication studies, model performance is worse with distractor squares, and line connectors improve performance only in the condition with squares. Rather than relying on abstract grouping principles, we show how the empirical results can be explained in terms of observers implementing a task-dependent grouping strategy that promotes overall performance.

尽管感知分组已被广泛研究，但其机制仍然知之甚少。我们提出了一个分组的神经模型，通过对其电路的自上而下控制，实现一个分组策略，包括连接策略（连接哪些元素）和选择策略（定义选择信号的时空属性以分割目标元素并便于识别）。我们将该模型应用于字母识别任务，该任务调查了统一连通性与邻近性和形状相似性的分组原则之间的关系。参与者报告了小圆圈是否形成了一个全局字母 E 或 H，这些圆圈可以通过一条线连接或嵌入一个正方形矩阵中。在模型中，该任务的一个好的分组策略包括在所有条件下连接圆形而不是正方形的连接策略和使用不同大小的选择信号的选择策略，具体取决于是否存在正方形。与在两个复制研究中得到验证的经验结果一致，模型性能在使用干扰方块的情况下更差，而线路连接器仅在使用方块的情况下才能提高性能。我们不依赖抽象的分组原则，而是展示了如何根据观察者实施促进整体绩效的任务相关分组策略来解释经验结果。这在两个复制研究中得到了验证，模型性能在使用干扰方块的情况下更差，而线路连接器仅在使用方块的情况下才能提高性能。我们不依赖抽象的分组原则，而是展示了如何根据观察者实施促进整体绩效的任务相关分组策略来解释经验结果。这在两个复制研究中得到了验证，模型性能在使用干扰方块的情况下更差，而线路连接器仅在使用方块的情况下才能提高性能。我们不依赖抽象的分组原则，而是展示了如何根据观察者实施促进整体绩效的任务相关分组策略来解释经验结果。