In traditional deep learning, information typically flows in one direction: sensory inputs are transformed into increasingly abstract representations through a feedforward chain. However, biological evidence and emerging computational theories suggest the brain operates differently. Instead of just filtering input, the brain acts as a proactive inference engine that constantly predicts its own sensory experiences.
This paradigm, known as Predictive Coding (PC), posits that the core objective of a neural system is to minimize the discrepancy between its internal expectations and the external world.
Perception as Statistical Inference
At its heart, predictive coding treats perception as a Bayesian "inverse problem." As Daniel Kersten and Pascal Mamassian argue in Object Perception as Bayesian Inference, the visual world is inherently ambiguous. A single 2D image on the retina could be caused by infinitely many 3D scenes. To resolve this, the brain combines noisy sensory data with "priors"—deep-seated knowledge about the structure of the environment.
In this framework, neural activity does not represent the stimulus itself, but rather a hypothesis about the stimulus. When a higher-level hypothesis successfully explains sensory data, it sends a signal downward to suppress the activity of lower-level neurons. This "silencing" mechanism, described by Daniel Kersten in Bayesian Models of Object Perception, ensures that the brain only spends energy processing the "residual error"—the parts of the signal it didn't expect.
The Mechanism of Predictions and Errors
A Predictive Coding Network (PCN) organizes neurons into a hierarchy where each layer serves a specific role in a bidirectional dialogue:
- Top-Down Predictions: Higher layers generate expectations about the activity in the layer below. These act as dynamic priors, biasing the lower layers toward a globally consistent interpretation.
- Bottom-Up Errors: Lower layers compare these predictions to their actual state. The difference—the prediction error—is the only information passed back up the hierarchy.
As William Lotter and Gabriel Kreiman demonstrated with PredNet, this architecture forces the network to learn the underlying structure of data. By trying to predict the next frame in a video, the model spontaneously learns object identity and motion without any human-provided labels. The higher layers are forced to model complex, long-term deviations that the lower layers cannot account for, leading to an efficient, unsupervised representation of the world.
Local Learning vs Global Backpropagation
One of the most significant advantages of predictive coding is its biological plausibility. Standard deep learning relies on backpropagation, which requires a global error signal to be sent backward through the entire network. This is widely considered biologically unrealistic because it requires neurons to have global knowledge of the network's weights.
In contrast, PCNs achieve global optimization through local interactions. As explored by Beren Millidge and colleagues in A Theoretical Framework for Inference and Learning in Predictive Coding Networks, these networks minimize a "global energy function" that is simply the sum of all local prediction errors.
Rafał Bogacz showed in An approximation of the error back-propagation algorithm in a predictive coding network that this process mathematically approximates backpropagation. By using local Hebbian plasticity—where synaptic changes depend only on the activity of connected neurons—PCNs can converge to the same solutions as artificial neural networks while remaining compatible with known brain anatomy.
Evidence from the Visual Stream
This theoretical framework is supported by direct neural observations. In Neural dynamics at successive stages of the ventral visual stream are consistent with hierarchical error signals, Elias Issa and James J. DiCarlo recorded from the macaque visual cortex. They found that neurons initially respond to a stimulus (the feedforward pass), but their activity is rapidly refined by feedback.
Specifically, they observed that neurons would "flip" their preferences shortly after the initial response—a dynamic behavior that feedforward models cannot explain, but which is a natural consequence of a system computing and resolving prediction errors. This suggests that the brain's visual hierarchy is not just a classifier, but a recurrent system designed to resolve sensory ambiguities through constant error correction.
Why It Works: Efficiency and Robustness
Predictive coding works because it optimizes for bandwidth and relevance. By suppressing predictable information, the system focuses its limited computational resources on novelty and surprise.
This makes the architecture inherently robust. As shown by Andrea Alamia and Rufin VanRullen in On the role of feedback in visual processing, when sensory input is noisy or "messy," the network naturally leans more heavily on its top-down expectations. The result is a system that doesn't just see what is there, but actively interprets the world based on what it knows should be there.