Fig. 1

Modulating conscious access using the Attentional Blink Paradigm. Due to copyright reasons, all photos except for the faces (which were photographed by one of the authors but have been anonymised) have been replaced by representational images. Eye regions are occluded above in the images to protect privacy but were not occluded in the experiment. a We presented a rapid serial visual presentation to participants, with two targets (T1 and T2) following each other within a stream of distractors. On the left, the second target (T2) is shown 200 ms after the first target (T1), and on the right, 800 ms after the T1. In every trial, participants had to detect and later recall both T1 and T2 targets. b We used a deep convolutional neural network (DCNN; yellow insert; 5 convolutional layers and 3 fully connected layers) to model the stimulus representational geometries (left) and predict our participants’ behaviour (right). The visual stimuli were fed into the DCNN, providing a hierarchical representation for each image. These unit activations were then analysed layer-by-layer and used to predict behaviour