Recurrent top-down synaptic connections at different spatial frequencies help disambiguate between dynamic emotions

AbstractThe coarse-to-fine hypothesis posits that, in the Human visual system, a coarse representation of visual information is propagated quickly through the retina to the cortex, whereas a finer, more detailed representation is propagated more slowly. In a previous study we showed that recurrent synaptic connections help predict low intensity EFEs. Furthermore, a feedback loop coming from coarser information processing is postulated to influence later processing of finer features. In this paper, we intend to examine the value of coarser information and recurrence in the processing of dynamic Emotional Facial Expressions (EFE). In a step forward in studying the importance of recurrent connectivity in the coarse-to-fine model, we tested its advantage for discriminating emotions for different spatial frequencies and facial expression intensities. Using Artificial Neural Networks, we modeled recurrent synaptic connections with a recurrent feedback loop. Using a Gabor filter bank, we computed different levels of spatial frequency features. Our results replicate the advantage of recurrence at first facial expression intensities. Our main finding is that the recurrent model is also better when predicting high spatial frequencies features. Additionally, mid-to-low spatial frequencies are more useful to the prediction of EFEs. We conclude that feature processing feedback has a significant effect in disambiguating facial expressions when information is particularly complex, i.e., at high spatial frequencies and low EFE intensities.

Return to previous page