Abstract
This review synthesizes advances in predictive processing within the sensory cortex. Predictive processing theorizes that the brain continuously predicts sensory inputs, refining neuronal responses by highlighting prediction errors. We identify key computational primitives, such as stimulus adaptation, dendritic computation, excitatory/inhibitory balance and hierarchical processing, as central to this framework. Our review highlights convergences, such as top-down inputs and inhibitory interneurons shaping mismatch signals, and divergences, including species-specific hierarchies and modality-dependent layer roles. To address these conflicts, we propose experiments in mice and primates using in-vivo two-photon imaging and electrophysiological recordings to test whether temporal, motor, and omission mismatch stimuli engage shared or distinct mechanisms. The resulting dataset, collected and shared via the OpenScope program, will enable model validation and community analysis, fostering iterative refinement and refutability to decode the neural circuits of predictive processing.
Introduction
Predictive processing (or predictive coding) suggests the brain uses an internal model to predict incoming sensory data. When predictions match reality, the signals are suppressed; when they do not, the brain generates a "prediction error" to update its model. While this is a powerful theoretical framework, exactly how neural circuits implement this at the hardware level (Marr’s Level 3) is a subject of ongoing debate.
The core challenge is that different laboratories utilize varied experimental paradigms to trigger prediction errors—such as visual oddballs, sensory-motor decoupling in virtual reality, or omitted stimuli. Because these studies often record from different brain areas and cell types using disparate setups, the field is characterized by conflicting findings. It remains unknown whether the brain employs a single, unified canonical circuit for all predictive processing or if different types of predictions rely on distinct neural mechanisms.
A Standardized Experimental Roadmap
To resolve these discrepancies, the authors propose a standardized suite of experiments to be conducted on mice via the Allen Institute's OpenScope program and on non-human primates through collaborating laboratories. This initiative aims to create a foundational open-source dataset to test competing models of brain computation.
The proposed methodology includes:
- Standardized Stimulus Sets: Four distinct "mismatch" paradigms tested on the same animals: standard visual oddballs, sensory-motor mismatches, sequence mismatches, and temporal duration mismatches.
- Dual Recording Modalities:
- In-vivo two-photon calcium imaging using transgenic mice to isolate specific, genetically defined cell types, particularly various inhibitory interneurons.
- Neuropixels electrophysiology to capture high-temporal-resolution spiking activity across multiple cortical layers and brain regions simultaneously.
- Unified Analysis Pipeline: A shared framework to evaluate whether error signals are additive, subtractive, or multiplicative, and to test specific computational models, such as dendritic versus cellular predictive coding.
Key Computational Primitives
The authors identify six mechanisms likely central to how the brain computes predictions:
- Stimulus adaptation
- Dendritic computation
- Cell-type specific computation
- Recurrent network dynamics
- Excitatory/Inhibitory (E/I) balance
- Hierarchical processing
The Interneuron Puzzle and Dendritic Hypothesis
Inhibitory interneurons—specifically PV (parvalbumin), SOM (somatostatin), and VIP types—play specialized roles in gating top-down predictions and bottom-up sensory data. Furthermore, the "Dendritic Hypothesis" suggests that apical dendrites of pyramidal neurons, which reach into superficial cortical layers, may be the primary sites where top-down predictions are compared against bottom-up sensory inputs, rather than the cell body.
Synthesis and Future Directions
The proposed OpenScope experiment is designed to test a fundamental hypothesis: does a unified predictive processing mechanism drive all mismatch processing in the cortex ($H_1$), or do mechanisms differ fundamentally depending on the type of prediction ($H_0$)?
Synthesis of existing literature reveals that while prediction error responses are often found in superficial layers (Layer 2/3) of the visual cortex, this segregation is less absolute in other modalities. Additionally, species differences between mice and primates necessitate cross-species validation. The field also remains divided on the communication of predictions: some evidence suggests specific brain rhythms (e.g., alpha/beta bands carrying predictions down the hierarchy, gamma bands carrying errors up), while other findings point to rapid, non-oscillatory transient spikes as the primary drivers.
Future work must also reconcile sensory predictive processing with reward prediction errors, which utilize different neural architectures and timescales. The resulting open-access dataset, shared via the DANDI archive in Neurodata Without Borders (NWB) format, will provide a vital resource for the global community to validate and refine predictive processing models.
Summary updated