Transformational Apparent Motion In A Recurrent Neural Network

Sudden onset of a shape that connects two disjoint static figures can induce perception of motion in the form of a continuous shape-change. We previously proposed that the direction of this Transformational Apparent Motion (TAM) is determined by shape correspondence independently of the features defining the shapes (Saleki et al., 2022). Here, using a Recurrent Neural Network (RNN, 128 hidden units), we explored the processing involved in the perception of this illusory motion. We trained an RNN using reinforcement learning framework on a perceptual decision-making task with TAM stimuli. The motion direction for each stimulus configuration was labeled to the left, to the right, or colliding in the middle according to the illusory motion direction reported in previous studies with human participants. The RNN was able to learn the task completely (100% accuracy on test dataset). Principle Component Analysis (PCA) revealed that network activity in different stimulus conditions diverged after the onset of the connecting stimulus on each trial (simulated with 4 time steps for the equivalent of 200 ms between the onset and response), and occupied different parts of the activity space by the end of trial. This was confirmed by representational similarity analysis, indicating that the representations of leftward and rightward motion had a greater distance to each other than to the motion toward the middle. Using PCA loadings, we found the most discriminative units in the network and observed different clusters based on characteristic response profiles. To evaluate feature independence, we tested the network that was trained on solid shapes on a separate dataset where stimuli were defined only by their outlines. The network achieved perfect performance in this task as well, showing similar characteristics in most of the analyses. Our findings provide insights into the underlying processes involved in perception of motion in TAM.