Different studies have shown the efficiency of a feed-forward neural network in categorizing basic emotional facial expressions. However, recent findings in psychology and cognitive neuroscience sug- gest that visual recognition is not a pure bottom-up process but likely involves top-down recurrent connectivity. In the present computational study, we compared the performances of a pure bottom-up neural network (a standard multi-layer perceptron, MLP) with a neural network involving recurrent top- down connections (a simple recurrent network, SRN) in the anticipation of emotional expressions. In two complementary simulations, results revealed that the SRN outperformed the MLP for ambiguous intensities in the temporal sequence, when the emotions were not fully depicted but when sufficient contextual information (related to previous time frames) was provided. Taken together, these results suggest that, despite the cost of recurrent connections in terms of energy and processing time for biological organisms, they can provide a substantial advantage for the fast recognition of uncertain visual signals.