Unconscious recognition

Current project — Unconscious visual disambiguation with prior information

I study how prior knowledge reshapes what the visual brain represents — and whether it has to reach conscious awareness to do so. Using fMRI, ambiguous two-tone “Mooney” images, and a binocular masking technique that pushes those images out of sight, I ask whether the brain still applies a learned interpretation to a stimulus that a person reports never having seen. It does: across five visual regions, from V1 to inferotemporal cortex, the neural pattern evoked by a suppressed ambiguous image shifts toward the pattern of its disambiguated counterpart — even though participants report no recognition at all, and even though the retinal input is identical before and after learning.

The question

Look at a two-tone Mooney image for the first time and you will probably see nothing but blotches. Then someone shows you the original photograph, and the blotches snap into a dalmatian, a face, a hand. Show you the same blotchy image again and you cannot un-see it. Nothing about the image changed; what changed is what you brought to it. Psychologists call this disambiguation, or one-shot perceptual learning, and it is one of the cleanest demonstrations that perception is a joint product of incoming sensory signal and stored knowledge.

Both the behavioural and the neural signatures are well documented. After disambiguation, people report recognizing the image far more often, and its neural representation drifts toward the representation of the clear photograph — despite the image on the retina being pixel-for-pixel the same.

What nobody had established is whether awareness of the ambiguous stimulus is required for any of this to happen. Prior work on unconscious vision has mostly asked whether invisible stimuli can be encoded — whether they can prime, enter memory, or drive learning. My question runs in the opposite direction: once a prior has already been acquired, can it be retrieved and applied to a stimulus that never becomes visible? And if so, how early in the visual hierarchy does that show up?

Why it matters

The answer bears directly on what conscious experience is for. Influential accounts tie consciousness to recurrent top-down signalling — the idea that experience arises precisely when feedback loops integrate predictions with sensory input. If the machinery that resolves visual ambiguity can run on a stimulus that produces no report and no recognition, then the processes that resolve ambiguity and the processes that generate a conscious report are at least partly separable. Either top-down prior application does not require awareness, or neural disambiguation is a more local, feedforward affair than the field has assumed. Both readings are consequential, and the paradigm lets me put a number on the effect rather than argue about it.

How I do it

The paradigm. Each image passes through three recognition stages: pre-disambiguation (Mooney image, never seen before), grayscale (the clear photograph that supplies the prior), and post-disambiguation (the identical Mooney image, now interpretable). Comparing the first and third stages holds the physical stimulus constant and varies only what the observer knows — which is exactly the manipulation the question needs.

Making the stimulus invisible. Every image is presented twice, once under normal binocular viewing and once under discontinuous flash suppression (dCFS): a high-contrast, 10 Hz coloured Mondrian animation goes to the dominant eye while the target goes to the other, and the target drops out of subjective awareness. I used the discontinuous variant, flashing the stimulus in short bursts, because it makes participants far less likely to break through the mask. Target opacity is set per participant with a two-stage staircase, so each person receives as much visual information as suppression allows — the conscious and unconscious conditions then differ only in what reaches the dominant eye.

Stimulus selection. From a 120-image library, a pilot study identified images that were reliably unrecognizable before disambiguation and reliably recognizable after. Ranking those by the size of the confidence jump left a final set with a mean increase of roughly 50 points on a 100-point scale — images that genuinely flip.

Imaging. 17 participants, 3T fMRI, seven runs structured so that a rolling subset of images is disambiguated in each run while others are held back, giving all three stages within a single session. Preprocessing runs through fMRIPrep; single-subject GLMs model each condition with motion parameters, drift terms and spike regressors, with head motion and temporal SNR checked before anything else is done.

Analysis: representational similarity. Univariate activation is the wrong tool here — the question is not how much a region responds but what it represents. So for each region of interest I build a representational dissimilarity matrix over all images and stages using correlation distance, then ask a targeted question of it: is a post-disambiguation image’s pattern more similar to its own grayscale counterpart than a pre-disambiguation image’s pattern is?

The critical methodological point is that raw similarity is contaminated. Repeated exposure, adaptation, rising confidence over the session, and generic semantic labelling would all inflate similarity between later stages regardless of which image is which. To isolate the genuinely stimulus-specific effect, I subtract from each matched (diagonal) pair the mean of that image’s similarity to the other eleven images at the same stage. What survives is the image-specific component and nothing else — which also disposes of the most obvious alternative explanation before it gets started.

Five regions — V1, V2, fusiform gyrus, inferotemporal cortex and MT — were fixed a priori from earlier work and defined from an independent meta-analytic atlas, so no voxel is selected using the data it is later tested on.

What I’ve found

No conscious disambiguation under suppression. Subjective identification under dCFS stayed at essentially zero and did not budge across recognition stages. Phenomenally, nothing happened.
Neural disambiguation happened anyway. In all five regions, the suppressed post-disambiguation image’s representation was significantly closer to its own grayscale counterpart than the pre-disambiguation image’s was — with large effect sizes, after the correction described above.
The conscious effect replicated, confirming that the intermittent presentation format did not break the phenomenon.
It also appears in a third, telling condition: trials where the image was fully visible but the participant reported being unable to identify it. So the effect tracks stimulus presentation, not subjective report.
The unconscious effect was never weaker than the conscious one in any region, and in MT it was reliably larger — which cuts against the common assumption that unconscious effects are just faded copies of conscious ones.

What I am careful not to claim

Awareness here was indexed by self-reported suppression breaks plus a binary recognition response, not by an objective, sensitivity-based measure — no forced-choice d′ near zero, no Bayes factors favouring the null. That is a real limit, and it means near-threshold perception on some trials cannot be excluded. So the honest description of the masked condition is greatly reduced stimulus visibility rather than certified unawareness, and the implications for theories of consciousness stay tentative until objective awareness controls are added. Suppression depth also varies with spatial frequency, and Mooney and grayscale images differ in exactly that respect.

Where it goes next

The obvious extension is to redo the visibility side of this properly: objective detection criteria, participant-level exclusion for above-chance performance, and ideally a presentation method that avoids interocular masking altogether. With awareness pinned down that tightly, the same representational analysis becomes a direct test between “top-down priors operate without awareness” and “neural disambiguation is more feedforward than we thought” — which is the question I actually want answered.

fMRI data collected at the National Taiwan University Imaging Center for Integrated Body, Mind, and Culture Research.