This paper provides a comprehensive synthesis of current advances in predictive processing theories, specifically focusing on the sensory cortex. Predictive processing posits that the brain continuously generates predictions of sensory inputs to optimize the processing of unexpected events, known as prediction errors. This research identifies key computational elements integral to this framework, such as stimulus adaptation, dendritic computation, excitatory/inhibitory balance, and hierarchical processing.
Key Insights and Numerical Highlights
The paper delineates the convergence and divergence of predictive processing mechanisms across species and sensory modalities. Notably, it highlights strong convergence in the role of top-down inputs and inhibitory interneurons shaping mismatch signals. Numerical emphasis is placed on the proportion of neurons responding to mismatches, such as the approximate 11% of excitatory neurons in layer 2/3 displaying mismatch responses in visual cortex paradigms. Contrastingly, it recognizes divergence in species-specific cortical hierarchy, with primates exhibiting deeper hierarchical levels compared to rodents, and modality-dependent roles of cortical layers.
Experimental proposals are suggested to resolve conflicts and knowledge gaps, primarily using in-vivo techniques such as two-photon imaging and electrophysiological recordings. The outcomes of these experiments, shared through the OpenScope program, aim to facilitate model validation and iterative refinement to decode the neural circuits involved in predictive processing.
Mechanistic Exploration
The paper describes a diverse array of error and mismatch types utilized in experimental settings (e.g., sensory mismatches, sensory-motor mismatches, and omission oddballs), each eliciting unique neuronal responses that may be supported by distinct biological mechanisms. For example, sensory-motor mismatches expose the brain's ability to integrate motor activities with corresponding sensory feedback, suggesting a role for top-down inputs from motor areas in shaping prediction errors.
At the cellular and network level, the role of different subtypes of neurons, particularly inhibitory interneurons such as PV, VIP, and SOM cells, is elucidated. These subtypes contribute to the computation and modulation of predictive signals, with VIP neurons involved in disinhibition and SOM neurons integrating lateral contextual inputs. Diverse responses in predictive processing can emerge from dendritic events, particularly within apical dendrites, which receive predictive signals from higher-level areas, modulating neuronal gain and encoding prediction errors.
Theoretical Modeling and Implications
The paper proposes theoretical modeling frameworks that simulate predictive processing using biologically plausible networks. Hierarchical models such as cellular and dendritic predictive coding propose varying pathways for information flow and integration across cortical areas. These models are grounded in recent findings of cortical connectivity patterns and interlaminar interactions.
Divergence between experimental evidence and theories is noted, particularly regarding hierarchical assumptions differing across species and the temporal dynamics of prediction errors. The paper advocates for integrative studies across modalities and species to refine existing models, suggesting these experiments could pave pathways for understanding predictive processing as a more generalized cortical computation.
Future Directions and Speculative Insights
The implications of this research extend to improving our understanding of sensory processing, neural plasticity, and cognitive functions. Future developments in artificial intelligence could leverage insights from predictive processing to enhance models that mimic brain-like computations. The iterative validation approach through collaborative datasets aims to refine theoretical frameworks continuously, laying grounds for more comprehensive models that could unify disparate findings in neuroscience related to predictive coding.
Conclusion
This paper provides a robust synthesis of the neural mechanisms governing predictive processing, underscoring both theoretical and experimental advancements. By proposing a unified experimental framework across modalities and species, and leveraging community collaboration, it sets the stage for resolving existing conflicts and advancing our understanding of predictive processing in sensory and cognitive neuroscience. The anticipated datasets promise to be instrumental for further research and model development, anchoring predictive processing as a central theme in understanding and simulating neural computations.