Neural implementation of categorization in a motion discrimination task

Human being can categorize one object into different classes depending on the reference used, a cognitive capacity, i.e., context-dependent categorization, which is fundamental in our daily life. In the present study, we explore one possible neural mechanism underlying a motion discrimination task, in which the neural system needs to judge whether a motion direction embedded in a random dot kinematogram is clockwise or anticlockwise with respect to a reference direction that varies over time. We construct a spiking-neuron network model to implement this task. The model consists of three parts: (1) a working memory circuit, which holds the information of the reference direction; (2) two information extraction circuits, referred to as clockwise-preferred circuit and anticlockwise-preferred circuit, respectively, which extract either the clockwise or anticlockwise information about the test direction; and (3) a decision-making circuit, which reads out the category decision. At the core of the network is the assumption of an asymmetric offset and rotational invariance of the connectivity profile. Our model successfully implements the context-dependent categorization of motion direction where the reference varies over time. And it reproduces the experimental results that with higher similarity between the reference and test direction or lower coherence level of the random dot kinematogram, the performance gets worse (lower accuracy and longer reaction time).