Figure 1
Experimental paradigm. Example of the sequence of events within an experimental trial. Participants were asked to discriminate the category of the masked image (animate v.s. inanimate) and then rate their visual awareness on a trial by trial basis.
Figure 2
Representational dissimilarity matrix of the hidden representations of the ResNet50 model fine-tuned by Caltech101 dataset. The images were resized to 128 × 128 × 3 and then passed through the model to obtain the feature representations. The RDM was computed by 1 – Pearson correlations of the feature representations. The first 48 items were animate and the last 48 items were inanimate.
Figure 3
Illustration of the pipeline used for encoding-based RSA across subjects and awareness states (i.e. from conscious to unconscious) or within the conscious or within the unconscious trials (not shown in the Figure). First, an encoding mode is trained to predict voxelwise responses based on the hidden layer representation of a convolutional neural network (CNN). The model is trained on the conscious trials of one participant and tested on the unconscious trials of a different participant in a standardized brain space. The predicted BOLD responses are then used to compute a predicted RDM across item pairs for a given searchlight sphere. The lower fields of computing the correlation between the predicted and true RDMs constitute the RSA.
Figure 4
Searchlight encoding-based RSA results in different cross-validation procedures. (a) within the conscious trials, (b) generalization from conscious to unconscious trials, and (c) within the unconscious trials. The Spearman’s rank correlation is computed between the predicted RDM based on the encoding model and the true RDM based on the BOLD responses of each moving searchlight sphere. Correlation coefficients were assigned to the center of the moving searchlight sphere. We show significant clusters that have been whole-brain corrected using TFCE (p < 0.05). The color bars show the range of correlation coefficients between 0 and 0.03. The left/right images of each panel display the left/right hemisphere. Corr. on the colorbar stands for correlation coefficient.
Figure 5
Searchlight decoding results in different cross-validation procedures. (a) within the conscious trials, (b) generalization from conscious to unconscious trials, and (c) within the unconscious trials. Decoding was conducted using a searchlight algorithm. ΔROC AUC on the colorbar represent the difference between the empirical ROC AUC (Receiver Operating Characteristic Area Under the Curve) and the theoretical chance-level score (0.5). The red color scale (0 to 0.05) corresponds to absolute differences from the chance level. A score of 0.05 indicates a 5% improvement over chance (i.e., ROC AUC = 0.55). the upper limit of the color bar is 0.05 (ROC AUC = 0.55). This cap was set to emphasize the range of biologically plausible effects, though individual voxels within significant clusters may exceed this value. The lower limit is 0 (ROC AUC = 0.5). ΔROC AUC scores were assigned to the center of the moving searchlight sphere. We show significant clusters that have been whole-brain corrected using TFCE (p < 0.05). The left/right images of each panel display the left/right hemisphere.
Figure 6
Results from the encoding-based searchlight RSA and decoding on the observers that showed null perceptual sensitivity. (a) Encoding-based RSA results, generalization from conscious trials to unconscious trials; (b) Decoding results, generalization from conscious trials to unconscious trials; (c) encoding-based RSA results within unconscious trials; and (d) decoding results within unconscious trials. ΔROC AUC on the colorbar represent the difference between the empirical ROC AUC (Receiver Operating Characteristic Area Under the Curve) and the theoretical chance-level score (0.5). Corr. on the colorbar stands for Spearman Rank correlation coefficient. The red regions show clusters where the Correlation coefficients or ROC AUC scores are significantly higher than chance level. Correlation coefficients or ROC AUC scores were assigned to the center of the moving searchlight sphere. We show significant clusters that have been whole-brain corrected using TFCE (p < 0.05). The color bars show the range of correlation coefficients or ROC AUC scores between 0 and 0.04. The left/right images of each panel display the left/right hemisphere.
Figure 7
Searchlight RSA results in different cross-validation procedures using VGG19 as the encoding model. (a) within the conscious trials, (b) generalization from conscious to unconscious trials, and (c) within the unconscious trials. Correlation coefficients were assigned to the center of the moving searchlight sphere. We show significant clusters that have been whole-brain corrected using TFCE (p < 0.05). The color bars show the range of correlation coefficients between 0 and 0.03. The left/right images of each panel display the left/right hemisphere.