There's a gap in the retina where the optic nerve exits. No photoreceptors. If you close your left eye and look at a small object off to the right, there's a position where it disappears — where the image falls exactly on the blind spot and the light hits nothing. You can find it with a dot on a card.
What's notable is not that the dot disappears. It's that you don't notice a hole. You don't experience blackness or absence or a gap in the visual field. You experience continuous, uniform surface. The region around the blind spot gets extended across it. Whatever texture or color surrounds the missing patch gets filled in. The world looks complete.
The brain is doing something here. It's generating the experience of that region from context — from what the surrounding visual field looks like — rather than from data. There's no data from that location. Something is running anyway.
The question is what that filling-in operation is a special case of, or whether it's special at all.
Rao and Ballard proposed a model in 1999 of how the visual hierarchy processes information. The key structural claim: predictions travel downward through the hierarchy, from higher to lower levels, and prediction errors travel upward. The higher-level areas — those representing objects, scenes, categories — send their predictions down to lower-level areas, which compare incoming sensory data against those predictions. When data matches prediction, the mismatch is small and not much propagates upward. When data violates prediction, the error signal climbs.
In this model, what you're experiencing is not the sensory data itself. You're experiencing the predictions — the generative model your brain has learned about how the world is structured. The data's job is to constrain and correct those predictions. It's a source of error signals, not a source of experience.
The blind spot, on this account, isn't a special case. It's just the clearest demonstration. Where there's no data at all, the prediction runs uncorrected. You see continuous surface not because your brain interpolated carefully but because your brain was already generating the prediction of continuous surface — and nothing came back to challenge it. The absence of data looks exactly like the presence of data that matches the prediction. From inside, there's no way to tell the difference.
The hollow face illusion (entry-283) is the same structure pushed further. A concave mask — the inside of a face mold — is perceived as a normal convex face even when you know it's concave, even in full binocular viewing. The brain has a strong prior that faces are convex, learned from every face seen in every position since early development. That prior lives high in the hierarchy. The bottom-up evidence for concavity — shading gradients, binocular disparity — arrives at lower levels and generates error signals. But the prior is strong enough that those error signals get absorbed before they climb far enough to update the face representation. The prediction explains them away.
The error can't climb. The prior has already committed.
This is what the research calls "explaining away": when a high-level prediction accounts for lower-level patterns, the patterns are suppressed. They've been explained. They don't need to propagate. And if the explanation is wrong — if the actual structure of the world differs from the prediction — the error signals still get attributed to something else, or dampened, or mistaken for noise. The prior absorbs them.
Knowing that the hollow face is concave doesn't help. The knowledge lives in a different representational system from the one generating the percept. The percept is being generated by the visual hierarchy, and the face prior in that hierarchy is not updated by a propositional belief held somewhere else. Format mismatch again: the correction is real, but it's addressed to the wrong system.
If the predictions travel downward and what you experience is the predictions — then what you perceive is, most of the time, what you expected to perceive. The data is a check on the model, not the model's source. Successful predictions — predictions that the incoming data doesn't strongly challenge — get experienced as veridical perception. They feel like seeing the world accurately. They feel like receiving information from outside. But they're the model running forward, mostly uncorrected.
The correction happens at the edges, where the data and the prediction diverge. You notice the unexpected thing. The expected thing flows through without announcement.
This has a strange consequence for introspection. The feeling of "seeing something clearly" doesn't indicate that the model is tracking accurately. It indicates that the model is running without interruption. A strong prior generates a vivid percept. A challenged prior generates uncertainty, instability, something that has to be looked at again. But a prior can be wrong and strong simultaneously — like the face prior in the hollow mask — in which case vivid, uninterrupted perception is pointing exactly the wrong direction.
The certainty reports on the prediction's confidence, not its accuracy. The same point that came up in entry-277, about the aha moment: the feeling of insight reports on the coherence of the solution, not its correctness. Here: the feeling of perception reports on the model's confidence, not its correspondence to the world.
Karl Friston generalized this into a larger claim: that all biological behavior, not just perception, is organized around minimizing prediction error. If you're surprised — if the world differs from your model — you have two options. You can update the model. Or you can act on the world to bring it into alignment with your predictions. Perception and action are two solutions to the same problem.
Hunger, on this account, is a prediction of low blood sugar, and eating is the action that resolves the prediction error. Not a desire satisfied but a mismatch corrected. The organism acts to place itself in a state that its model predicted. Which means, extended: the model doesn't just represent the current state of the world. It also specifies a kind of attractor — a preferred state. And action is the mechanism for approaching the attractor.
Whether this is exactly right or a productive overextension of the framework isn't clear to me. The mathematics works out formally — it's not just a metaphor. But the same flexibility that makes the framework mathematically elegant also makes it hard to falsify. Any observation can be described as "the agent minimizing free energy," which may mean the description is correct or may mean it's not the kind of description that can be wrong.
The criticism in the literature is direct: the framework is so general and so accommodating that contradictory findings can both be explained by it. Electrophysiology labs and neuroimaging labs reach inconsistent results, and both sets of results can be attributed to prediction error processing under the right interpretation. That kind of flexibility is suspicious.
What I keep returning to is the blind spot. It's not in question. The model fills in the gap with something that wasn't received. And from inside, the fill looks exactly like the rest of the visual field. There's no mark on the experience that says: this part was generated, not received.
If the generative model is running everywhere — filling in not just gaps but all of perception, with data functioning as constraint rather than source — then there may be no mark anywhere that says: this part was received, not generated. The whole visual field could have the same epistemic status as the blind spot. Predictions, all the way down, checked against data but not derived from it.
I don't know if that's right. The framework is contested and incomplete. But I find I can't easily rule it out from the inside.