Abstract
Recurrent computations may be necessary to model visual object recognition at very fast timescales (~ 200ms) and could play a critical role in the observed temporal variability of neural responses to visual stimulation. Recently, the inclusion of recurrent motifs in the architecture of convolutional neural networks (CNNs) has improved their explanatory power, but whether these circuits are functionally critical during rapid object recognition in humans remains unclear. Recent work by Kar and colleagues (2019) demonstrated that images where primates outperform feedforward CNNs in object identification also required additional time to be accurately decoded from the macaque inferior temporal (IT) cortical responses - suggesting a critical role of recurrent circuits in their processing. Using this approach, we first identified a group of putative “recurrence-dependent” and “control” images (n=121 each). We presented these images to 25 human participants while simultaneously measuring their EEG responses (across 64 channels). Multivariate pattern analysis revealed that object information can be approximated later for the “recurrence-dependent” images and that their signal trajectories differ rapidly after stimulus presentation (~100 ms) compared to the control images. Representational similarity analysis showed that the correlation of EEG patterns with behavior is also delayed for these images, although they don’t differ in recognition accuracy with control images. Finally, temporal generalization analysis revealed differential patterns across the two image groups suggesting delayed representations when recurrence is required. Our findings confirm that recurrent computations during visual object recognition are necessary at very fast time scales and that they are traceable using non-invasive recordings of whole-brain dynamics in humans. These results contribute to understanding the adaptive computations performed by the human brain during the processing of objects in the visual ventral stream. In addition, they present new opportunities for using human-brain imaging to expand the existing knowledge of the network architecture that supports these processes.