How our brains, a 3-pound mass of tissue encased in a bony skull, creating perception from sensations has been an age-old mystery. Abundant evidence and decades of ongoing research show that the brain cannot simply assemble sensory information, as if it were putting together a jigsaw puzzle, to perceive its surroundings. This is shaped by the fact that the brain can create a scene based on light entering our eyes, even when the incoming information is noisy and ambiguous.
As a result, many neuroscientists are turning to the brain as a “prediction machine”. Through predictive processing, the brain uses its previous knowledge of the world to infer or hypothesize the causes of incoming sensory information. Those assumptions – not the sensory inputs themselves – give rise to perception in our mind’s eye. The more ambiguous the input, the greater the dependence on prior knowledge.
“The beauty of predictive processing frameworks [is] that it is really big – sometimes critics may say too loud – the ability to explain so many different phenomena in many different systems,” Floris de Lange, a neuroscientist at Radboud University’s Predictive Brain Laboratory in the Netherlands.
However, the growing neuroscientific evidence for this idea is largely circumstantial and open to alternative explanations. “If you look at cognitive neuroscience and neuroimaging in humans, [there’s] lots of evidence — but indirect, super-hidden evidence,” said Tim Kietzmann of Radboud University, whose research lies in the interdisciplinary fields of machine learning and neuroscience.
Therefore, researchers are switch to computational model to understand and test the idea of the predictive brain. Computational neuroscientists have built artificial neural networks that, with designs inspired by the behavior of biological neurons, learn to make predictions about incoming information. These models show some strange abilities that seem to mimic those of the real brain. Some experiments with these models even suggest that brains must have evolved as predictive machines to satisfy energy constraints.
And as computational models evolve, neuroscientists who study living animals increasingly believe that the brain learns to infer the cause of sensory inputs. While the exact details of how the brain does this are still vague, the broad strokes are becoming clearer.
Unconscious reference in perception
Predictive processing may seem like a counterintuitive complex cognitive mechanism at first, but there is a long history of scientists using it for other seemingly desirable explanations. Even a thousand years ago, the Muslim Arab astronomer and mathematician Hasan Ibn Al-Haytham highlighted a form of it in Optical Book to explain different aspects of vision. This idea gathered strength in the 1860s, when German physicist and physician Hermann von Helmholtz argued that the brain infers external causes of sensory inputs instead of constructing its own perceptions. “bottom-up” from those inputs.
Helmholtz interpreted this concept of “unconscious inference” to explain dual-stable or multi-stable perception, where an image can be perceived in more than one way. This happens, for example, with the famous ambiguity we might perceive as a duck or a rabbit: Our perception is constantly flipping between the two animal images. In such cases, Helmholtz asserts that perception must be the result of an unconscious process of top-down inference about the cause of the sensory data because the images formed on the retina do not change.
Throughout the 20th century, cognitive psychologists continued to make the case that perception is an active constructive process based on both bottom-up and top-down conceptual input. The effort culminated in an influential 1980 article, “Hypothetical perception, “The end Richard Langton Gregory, argues that perceptual delusions are essentially false assumptions by the brain about the cause of sensory impressions. Meanwhile, computer vision scientists have stumbled in an attempt to use bottom-up reconstruction to allow computers to see without an internal “reproductive” model for reference.
