Abstract
Predictive coding (PC) is a unifying framework for understanding perception, action, and neocortical organization. In this framework, different areas of the neocortex implement a hierarchical generative model of the world learned from sensory inputs. Cortical circuits are hypothesized to perform Bayesian inference based on this generative model: top-down feedback connections convey predictions of lower-level activities, while bottom-up feedforward connections convey the errors between these predictions and actual sensory inputs. By minimizing prediction errors, predictive coding provides a functional explanation for a wide range of neural responses and fundamental aspects of brain organization.
Introduction: Perception as Inference
Traditional "bottom-up" models of vision often struggle to explain why perception is so heavily influenced by context and prior expectations. Perception is fundamentally an "ill-posed" inference problem; a single 2D retinal image could be caused by an infinite number of 3D configurations. The brain resolves this ambiguity by using an internal generative model to "guess" the causes of sensory inputs.
As illustrated in Figure 1, our perception of a footprint as either an indentation or a mound depends on where we assume the light is coming from. Predictive coding treats these assumptions as "priors" that constrain the interpretation of sensory data.
The Rao-Ballard Hierarchical Model
The core of this theory is the Rao-Ballard model, a hierarchical framework where each level of the cortical hierarchy attempts to predict the activity of the level below it. Mathematically, the input $I$ is modeled as $I = f(Ur) + n$, where $r$ is the hidden state (the "cause"), $U$ represents learned synaptic weights (basis functions), and $f$ is a non-linear activation function.
By applying gradient descent to a cost function that minimizes reconstruction error and enforces priors (like sparsity), we derive update equations that mirror biological neural networks:
- Feedback connections carry predictions downward.
- Feedforward connections carry the residual error upward to update the internal state.
Key Insights and Biological Mapping
Predictive coding provides a functional logic for the brain's physical structure:
- Surprise as the Signal: In a predictable environment, a PC-based brain is efficient and relatively silent. Neural activity in the feedforward path represents "surprise" (the difference between expectation and reality), which is more information-dense than repeating known data.
- Laminar Specificity: The framework suggests that Layer 2/3 neurons act as "error detectors" (sending signals up), while Layer 5/6 neurons maintain the "state estimate" and send predictions down.
- Attention as Precision: Attention acts as a gain control on prediction error. If a signal is noisy, the brain lowers the "precision" weight of that error, effectively deciding to ignore the mismatch.
Results and Validation
PC models have successfully replicated several observed neural phenomena. When trained on raw natural image patches, these models spontaneously develop Gabor-filter-like receptive fields identical to those found in the primary visual cortex (V1).
Furthermore, the model explains "extra-classical" effects like endstopping, where a neuron's response to a bar of light decreases if the bar becomes too long. In PC, a longer bar is more predictable by higher-level neurons, which send a stronger inhibitory prediction down, suppressing the error signal. Experimental evidence supports this; for example, studies in the mouse visual cortex show that roughly 13% of neurons respond specifically to "mismatches" between expected and actual visual flow.
Limitations and Future Work
While powerful, current PC models often use MAP (Maximum A Posteriori) estimation, providing a single "best guess." A key frontier is moving toward "Variational" PC that handles full probability distributions to represent uncertainty. Additionally, while the framework explains spatial processing well, how the hierarchy segments information across the massive range of biological timescales remains an open question. Finally, the "Dark Room Paradox"—the idea that error minimization should lead us to seek sensory deprivation—is addressed by Active Inference, which posits that our biological priors include the expectation that we are active, surviving organisms, forcing us to seek out expected sensory states.
Summary updated