Brain and DCNN representational geometries predict variability in conscious access

Abstract

When two targets (T1 and T2) are presented in a rapidly sequentially-presented stream of distractors, subjects often show a clear deficiency to report T2 when presented 200-500 ms after T1. This effect is known as the Attentional Blink (AB). Using the AB as a method to quantify the probability of conscious access, we investigate why some images seem to rise to consciousness more readily. By defining the representational relationships between images using fMRI and CNNs, we show that images that are distinct in high-level representations are more resilient to the AB effect, while low-level similarity to other images increase the probability of conscious access. These results were replicated using representational geometries derived from both functional Magnetic Resonance Imaging (fMRI) and Convolutional Neural Network (CNN). This provides additional parallels between the hierarchical complexity of CNNs trained on object classification and the human visual ventral stream, with CNN and brain representations predicting behaviour in a similar way.