Abstract:

If subjective experience is a property of neural activity in the brain then differences in subjective state must map onto differences in neural activity in the brain. The brain responds to stimuli that are not consciously experienced, and these neural responses are both reliable (i.e. similar across different presentations of the same stimulus) and differentiated (i.e. distinct from Reponses provoked by other stimuli). What distinguishes the brain's response to stimuli that are consciously experienced from its response to those that are not? The aim of the present work was to examine the neural responses to sensory stimulation with and without subjective experience of the stimulus object, and to identify those differences that are independent of the particular anatomical locus or absolute intensity of the neural response.

While locus and intensity of response likely vary depending upon whether or not a stimulus is experienced subjectively, we predicted, based on the theory of Marcel Kinsbourne, that there would also be a qualitative difference in the pattern of cortical activation, reflected indirectly in the stability and reliability of the BOLD activation vector, both spatially and in terms of its angle in a high-dimensional space. We also predicted an abrupt non-linear transition between conscious and non-conscious representations, rather than a graded change, based on prior studies and our own behavioral data.

Dichoptic color masking was used to parametrically vary the visibility of face and house stimuli from just above the limen (near-perfect identification) to just below (chance-level identification). Post-decision wagering was used to infer the presence or absence of subjective experience at a given level of color contrast: after identifying the category of the stimulus (face or house, guessing if necessary) subjects wagered (high or low) on the accuracy of their perceptual decision. A lack of correspondence between high wagers and correct perceptual decisions (measured using d-prime) was taken to infer a lack of subjective experience.

In behavioral experiments we found that accuracy remained surprisingly high (~70% correct) even when wagering d-prime had reached zero, suggesting non-conscious processing of the stimuli. Eye-tracking data revealed that sub-saccadic eye movements may mediate this non-conscious effect. When stimuli were rendered difficult to discern based on low-level information (contour differences), we predicted that accuracy would descend to chance with decreasing color contrast, and then rebound to above-chance at even lower levels of color contrast, consistent with an abrupt non-linear transition. Our prediction of a "Luke Skywalker effect" was confirmed.

A slightly modified version of the same experiment was carried out in the MRI scanner. Event-related functional data were analyzed using a Gaussian na?ve-Bayes classifier (in collaboration with Francisco Pereira), trained to classify the stimulus category (face versus house) independently at each level of color contrast, restricted to voxels in the temporal lobes. The accuracy of the classifier was significantly above chance at both the highest level of color contrast ("seen"; subjects near ceiling on the identification task) and the lowest ("unseen"; subjects at chance), and the difference in accuracy between these two was not significant (63% and 58% correct respectively). When the classifier was trained on "seen" patterns and tested on "unseen", and vice-versa, the accuracy of the classifier fell to chance. These two results suggest that non-conscious representations are not merely noisy versions of conscious representations, but that the two types of representations are qualitatively different. The subsets of voxels selected as informative by the classifier revealed a striking difference between conscious and non-conscious patterns: the spatial distribution of informative voxels was significantly more reliable across exemplars (within subject) for conscious versus non-conscious patterns, with a prominent focus in the right ventral temporal lobe (across subjects). The spatial distribution of information in non-conscious patterns was highly variable from one exemplar to the next, suggesting that the information in non-conscious patterns is "disintegrated" and less coherent - consistent with our prediction.