Perception feels effortless. We open our eyes, and the world simply appears. But this apparent simplicity hides an extraordinary feat. The brain has no direct access to the world, instead it only receives an endless cascade of electrical signals from the eyes. These signals are noisy and ambiguous, yet within a fraction of a second, we can recognize faces, read emotions, and navigate our environment. How does the brain achieve this?
The brain is not a camera passively recording the world, instead it is an active prediction machine. It continuously generates predictions about what it believes to perceive next. When predictions match reality, processing runs smoothly. However, when something unexpected happens a large prediction error ensues, which in turn is used to update the brains believes about the world, thereby minimizing future prediction errors. This mechanism, known as predictive processing, is thought to be a fundamental organizing principle of the brain, involved in perception, attention, learning, etc.
But a key question has remained unanswered: What exactly does the brain predict? Vision is a hierarchy of features, from raw contrasts and edges to shapes and textures, to recognizable objects and their meaning. When you expect to see a dog, are you predicting the precise pattern of pixels, the characteristic shape of fur and ears, or the concept "dog"?
In a new study, now published in the journal iScience, researchers at the Mind, Brain and Behavior Research Center (CIMCYC) at the University of Granada (UGR) set out to answer this question.
The experiment: Predicted and surprising images
The research team (David Richter, Paula Pena, and María Ruz) designed an experiment in which participants viewed sequences of object images while their brain activity was recorded using electroencephalography (EEG). EEG captures the brain's electrical signals with millisecond precision. Each image was preceded by a cue that predicted which object would most likely appear. This allowed the researchers to measure how much each unexpected image surprised the brain.
Key to the analysis was a deep neural network, an AI used to recognize objects in photos. Using this deep neural network, the researchers could quantify surprise at different levels of visual abstraction. Early layers of the network capture low-level features (edges and contrasts) and late layers represent higher-level visual structure (e.g. what object a scene contains). This enabled the independent measure of low-level and high-level surprise elicited by each image.
The results: High-level predictions drive visual prediction errors
Neural responses starting around 190 milliseconds after an image appeared were selectively amplified when the object was more surprising in terms of high-level visual features. Low-level surprise had no detectable effect, even though the images were predictable all the way down to the pixel-level.
In other words, the brain did not appear to bother predicting the precise low-level features of upcoming images. Instead, it predicts what kind of thing is coming. Like a reader who anticipates the meaning of the next word rather than the exact shape of each letter, the brain's predictions may predominantly operate at a level of abstraction that is most useful for perception and behaviour.
Why does this matter?
This work advances our understanding of how the brain makes sense of the complex world around us. Specifically, the brain does not appear to laboriously check every minor visual detail but instead maintains and flags expectation violations at a higher level of visual abstraction. Understanding the contents of prediction, and not just that the brain predicts, but what it predicts, is essential for any theory of perception, attention, and learning.
Reference
Richter, D., Pena, P., & Ruz, M. (2025). Rapid computation of high-level visual surprise. iScience. https://doi.org/10.1016/j.isci.2025.114121
Contact
Maria Ruz (@email)