Random forest classification results:A regularized random forest (Scikit-learn implementation). The implementation utilized the cross-entropy as an objective function with 500 decision trees. This number of decision trees was used as part of the regularization in order to avoid overfitting and also to approximate a simpler model which would enable comparison with the results from the logistic regression. Tree-based models are composed of nodes with each representing a level of depth in the model resulting from binary decision nodes where each node compares one feature value of the samples to a threshold. The maximum depth of each decision tree classifier was set to 1 so that only 1 best feature would be eventually used to make the decision. This means that during the training, with bootstrapped samples and features, the individual decision tree classifier ranks the importance of the feature by dropping one of the features and estimate a new decoding score after the dropping. The more loss in the decoding score compared to the original score, the more important the given feature was. In other words, we aimed to estimate the best feature among all features used (i.g confidence, awareness, and correctness). The majority rule was then applied to estimate the feature importance across all decision tree classifiers.Note the sum of feature importance of all features add up to 1 in the implementation of Scikit-learn: ( https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html#sklearn . ensemble.RandomForestClassifier.feature_importances_ ), and so this will be the case across all trials back. Hence, the main effect of the time window in the ANOVA is not meaningful because the average feature importance at each time window would be exactly the same (i.e. 0.33) and consequential it is not reported.

A framework to pinpoint the scope of unconscious processing is critical to improve our models of visual consciousness. Previous research observed brain signatures of unconscious processing in visual cortex but these were not reliably identified. Further, whether unconscious content is represented in high-level stages of the ventral visual stream and linked parieto-frontal areas remains unknown. Using a within-subject, high-precision fMRI approach, we show that unconscious contents can be decoded from multivoxel patterns that are highly distributed alongside the ventral visual pathway and also involving parieto-frontal substrates. Classifiers trained with multivoxel patterns of conscious items generalised to predict the unconscious counterparts, indicating that their neural representations overlap. These findings suggest revisions to models of consciousness such as the neuronal global workspace. We then provide a computational simulation of visual processing/representation without perceptual sensitivity by using deep neural networks performing a similar visual task. The work provides a framework for pinpointing the representation of unconscious knowledge across different task domains.

No abstract

Our perceptual system appears hardwired to exploit regularities of input features across space and time in seemingly stable environments. This can lead to serial dependence e ects whereby recent perceptual representations bias current perception. Serial dependence has also been demonstrated for more abstract representations such as perceptual con dence. Here we ask whether temporal patterns in the generation of con dence judgments across trials generalize across observers and di erent cognitive domains. Data from the Con dence Database across perceptual, memory, and cognitive paradigms was re-analyzed. Machine learning classi ers were used to predict the con dence on the current trial based on the history of con dence judgments on the previous trials. Cross-observer and cross-domain decoding results showed that a model trained to predict con dence in the perceptual domain generalized across observers to predict con dence across the di erent cognitive domains. The recent history of con dence was the most critical factor. The history of accuracy or type-1 reaction time alone, or in combination with con dence, did not improve the prediction of the current con dence. We also observed that con dence predictions generalized across correct and incorrect trials, indicating that serial dependence e ects in con dence generation are uncoupled to metacognition (i.e. how we evaluate the precision of our own behavior). We discuss the rami cations of these ndings for the ongoing debate on domain-generality vs. speci city of metacognition.

The development of novel frameworks to understand the properties of unconscious representations and how they differ from the conscious counterparts may be critical to make progress in the neuroscience of vision consciousness. Here we re-analysed data from a within-subject, high-precision, highly-sampled fMRI study (N=7) coupled with model-based representational similarity analysis (RSA) in order to provide an information-based approach to study the representation of conscious and unconscious visual contents The standard whole-brain searchlight RSA revealed that the hidden representations of convolutional neural network models explained brain activity patterns in response to unconscious contents in the ventral visual pathway in the majority of the observers, particularly for models that ranked high in explaining the variance of the visual cortex (i.e., VGGNet and ResNet50). Also five of seven subjects showed brain activity patterns that correlated with the model in frontoparietal areas in the unconscious trials. However, the results of an encoding-based RSA analyses in the unconscious condition were mixed and somehow difficult to interpret, including negative correlations between the representations of the computer vision models and the brain activity in frontal areas in a substantial amount of the observers.