Predictive Coding Theory

Human perception is inherently distinctive; our remarkable perspectives are synthesised from a variety of lenses, ranging from social norms to memory and emotion. While having a unique perception may allow us to have diverging interpretations of the same stimuli, it can cause some realities to become more filtered than others. This process has been modelled through a framework “Predictive Coding Theory”, which is a theory that redefines how we understand perception itself. But what exactly does predictive coding actually propose?

Assumptions

Predictive Coding Theory (PCT) posits that consciousness is generated based on the brain’s continuous prediction and updating of its internal model of the world. In other words, our perception (or reality) is based on the brain’s predictions, actively anticipating upcoming stimuli and interpreting it instead of passively registering information.

The very goal of PCT is to minimise prediction errors — discrepancies between what the individual expects versus what actually happens in reality. This drive to actively reduce prediction errors is known as active inference, which is the brain’s continuous ability to update or correct predictions based on the sensory data received from our environment. 

According to Caucheteux, Gramfort & King (2023), there are four key assumptions that help explain how the brain processes information: 

1. Efficient neural coding: Like all biological systems, the brain has evolved to be as efficient as possible given the array of different information it receives on a daily basis. Having an efficient neural coding system allows the brain to prioritise relevant signals while suppressing redundant or expected input, conserving cognitive resources. 

2. Top-down predictions: A method where higher brain regions anticipate sensory input and relay predictions downward, shaping perception ahead of sensation — this will be explained in the next section in further detail. 

3. Minimising prediction error: As mentioned previously, minimising prediction errors is central as to to how the brain updates its internal models.  

4. Cognition as a form of probabilistic inference that uses generative models: That was one lengthy sentence, but in simpler terms, it means that the brain is constantly guessing what’s out there based on past experience, using its internal models to simulate possible realities and refine them as new data comes in. 

These assumptions are central for understanding how the brain implements predictive coding into practice in the following section.

Hierarchal Predictive Coding Model 

Predictions are generated in a hierarchical fashion, mainly divided into higher and lower-level processing. It’s theorised that all areas of the brain are responsible for generating predictions. 

When Rao & Ballard first formalised PCT into a general model for the cortex in 1999, they posited that the brain is categorised in layers, and each layer consisted of different parts of the brain responsible for specific jobs: 

#1 Higher-level areas: While all areas in this model are responsible for generating predictions, the high-level areas are the primary generators of predictions who initiate and shape them. This kind of processing involves abstract and conceptual functions. As the brain’s “supervisors”, they only need to intervene and send predictions downward when the sensory data is ambiguous, unexpected or contextually complex. 

• Prefrontal cortex: Generates goal-directed predictions, which shapes intentions and contextual outcomes. 

• Temporal & parietal cortices: Responsible for object identification and recognition, creating semantic meaning and managing spatial context. 

• Limbic system: Encodes emotional relevance, reward expectations, and interoceptive states (ex: hunger, pain). 

#2 Lower-level areas: precise and data-driven, focusing on raw sensory input. These regions compare incoming stimuli against predictions sent from higher-level areas. When there’s a mismatch between the prediction and reality, they generate prediction errors, which act as signals that are sent upward to help refine the brain’s internal models.

• Primary Sensory Cortices: Predicts basic features like edges, tones, motion, or texture.

• Thalamus & Brainstem: Anticipate timing and intensity of raw sensory input.

#3 Mid-level areas: integrative and semantic, acting as a translator between the higher-level and lower-level areas. These regions are key to refining and contextualising information in both directions. 

• Posterior Parietal Cortex: Supports spatial attention and coordinates between multiple sensory modalities. 

• Inferior Temporal Cortex: Contributes to object identity and generates both categorical and visual predictions. 

• Temporoparietal Junction: Plays a key role in social inference and causal reasoning. 

• Superior Temporal Sulcus:  Integrates dynamic feature integration (ex: biological motion, face processing), enabling nuanced perception of social and environmental cues.

When a prediction from a higher-level brain area matches the actual sensory input, the lower-level areas don’t need to engage in further processing—the input is considered “explained.” 

In this context, “explained” means that the brain has successfully accounted for the incoming information using its internal model, leaving little need for additional analysis. As a result, neural activity in those lower regions is actively suppressed through feedback from higher levels. This phenomenon is known as explaining away, where predicted input is effectively ignored because it aligns with expectation and requires no further cognitive effort.

How the Brain Predicts

As stated previously, the very goal of PCT is to minimise prediction errors. As such, when the brain processes incoming predictions, depending on which region it belongs to, it uses one of the following processing styles: top-down processing and bottom-up processing.  

Top-Down Predictions

Bottom-Up Predictions

Used by high-level regions The high-level regions use past experience, knowledge, and context to generate predictions.  They transmit these predictions downward towards the low and mid-level regions to shape how incoming data is processed.   Creates a clear direction for processing and enables fast decision-making. Perception becomes quicker and more efficient by anticipating incoming input. Processing flows from general concepts to specific details. This approach is prone to bias and can lead to perceptual illusions.

Used by low-level regions The low-level regions compare actual input to the high-level region’s predictions. If there’s a mismatch, they send error signals back up to higher regions. Processing moves from specific sensory details to broader interpretations. This method can be slower and more difficult to coordinate. Perception is more accurate and flexible, adapting to unexpected input. The senses provide raw data (ex: colours, sounds, textures) for evaluation. High-level regions update their predictions based on these error signals.

Evaluation of PCT

Like all theories, PCT is not an all-you-can-fit solution for explaining perception:

Strengths

Weaknesses

Attempts to be comprehensive: aiming to account for both low-level perceptual processes and high-level cognitive functions across domains.  Unifying framework: offers single computational principle (minimising prediction errors) to explain cognitive domains. Biologically grounding: matches with current cortical architecture (ex: hierarchical structure, feedback and feedforward loops).

Lacks sufficient explanations: too broad, was not backed up by enough empirical evidence.  Lack of consensus: there is zero consensus on:  how prediction and error neurons are organised how prediction errors are computed biologically the role of attention within this framework  how the brain weighs prediction errors  Alternative explanations: Findings can also be explained by simpler alternative models (ex: Bayesian Inference Frameworks, Classical Feedforward Models)

References

Caucheteux, C., Gramfort, A. and King, J.-R. (2023). Evidence of a predictive coding hierarchy in the human brain listening to speech. Nature Human Behaviour, [online] 7, pp.1–12. doi:https://doi.org/10.1038/s41562-022-01516-2.d

en Ouden, H.E.M., Daunizeau, J., Roiser, J., Friston, K.J. and Stephan, K.E. (2010). Striatal Prediction Error Modulates Cortical Coupling. Journal of Neuroscience, [online] 30(9), pp.3210–3219. doi:https://doi.org/10.1523/jneurosci.4458-09.2010.