Recurrent Pattern Completion Drives the Neocortical Representation of Sensory Inference

Academic Background: Exploring Perceptual Inference and Neural Mechanisms

In everyday life, our sensory systems often encounter incomplete or ambiguous information. For instance, when objects are occluded, the brain relies on prior experience and expectations to infer the whole. This inferential ability is not only a core function of the human visual system, but is also shared by other animals, including primates, rodents, fish, and even insects. Sensory inference not only enables us to identify edges and shapes that do not physically exist, such as the famous Kanizsa triangle illusion, where observers perceive a white triangle that is not actually present. This “illusory contour (IC)” phenomenon essentially reflects a high-level process of perceptual inference, the neural mechanisms of which have long remained poorly understood.

Previous studies in primates and humans have revealed that a subset of neurons in the primary visual cortex (V1) react to illusory contours as if a real edge is present. In higher visual areas, the gestalt perception that “the whole is greater than the sum of its parts” is even more pronounced for illusory contours. These neuronal responses are considered the low-level neural expressions of advanced perceptual inference. However, how illusory contours are encoded in cortical networks, how real edges and inferred edges are distinguished, and the neuronal circuit organization underlying these processes are still subject to controversy and unknowns.

This study addresses these scientific issues by systematically unveiling the neural coding and circuit mechanisms of neocortical sensory inference through high-throughput neural recordings and precise optogenetic manipulation, with a focus on the role of pattern completion in inference processes.

Source and Authorship

The article, entitled “Recurrent pattern completion drives the neocortical representation of sensory inference,” is authored by Hyeyoung Shin (corresponding author) and a team from University of California, Berkeley; Seoul National University; and the Allen Institute Neural Dynamics Program, among other internationally renowned research institutions. The paper was published in the prestigious journal “Nature Neuroscience” in November 2025, DOI: https://doi.org/10.1038/s41593-025-02055-5.

Overall Research Workflow & Technical Highlights

This study utilizes mice as experimental subjects and integrates cutting-edge neuroscience methods, including multi-channel Neuropixels (high-density electrophysiological probe) recording, two-photon (2p) calcium imaging, 2p holographic optogenetics, and a self-developed 2p holographic mesoscope (large field-of-view cellular resolution imaging), to systematically analyze the multilayered visual cortical representation mechanisms of illusory contours. The research can be summarized into several main stages:

I. Discovery of Illusory Contour-Encoding Neurons (IC-encoders) and Their Functional Differentiation

1. Experimental Subjects and Group Design: Subjects included mature C57/B6 mice and several transgenic backgrounds, such as camkii-tta;teto-gcamp6s mice (for V1L2/3 imaging), scnn1a-tg3-cre;ai162 mice (for V1L4 imaging), and SST and PV-Cre backgrounds (Neuropixels experiments). Each experiment involved robust sample sizes (e.g., neural recording experiments typically included 12 mice per group, and imaging experiments involved hundreds to thousands of neurons).

2. Innovative Visual Stimulus Protocols: To distinguish neuronal responses to global illusory contours from local inducing segments, the authors meticulously arranged stimulus images—including classic Kanizsa triangles, real edges (IRE images), illusory contours (IC images), newly designed L-shaped combinations (LC images), and TRE (real edge probe) and XRE image series. Through these composites, they accurately identified neurons selectively responsive to the global IC but not to local segments—i.e., “IC-encoders,” and neurons primarily responsive to individual segments, called “segment responders.”

3. Neural Activity Recording and Analysis Methods:

   • Multi-channel Neuropixels Probes: Simultaneous insertion into six visual cortical areas (V1, LM, AL, RL, AM, PM) of a single mouse, recording hundreds to over a thousand units.
   • Two-photon Calcium Imaging and the Self-developed Mesoscope: Large field of view, high spatiotemporal resolution observation of activity dynamics across thousands of neurons within the same cortical region.
   • Modeling Image Stimulus-Neuron Response Relationships: Use of SVM (support vector machine) and other linear classifiers to train mappings between neural activity and stimulus types, and further test inference ability (inference decoding).

4. Innovative Data Analysis Algorithms: The authors developed an inference decoding approach to differentiate neurons that genuinely encode illusory contour inference and conducted hierarchical comparisons across cortical regions.

II. Cortical Hierarchy of Illusory Contour Inference and Functional Partitioning of Neuron Subgroups

1. Hierarchical Regional Localization and Functional Identification: Multi-area neural recording and imaging indicated that the primary region driving IC inference is the L2/3 layer of V1, while V1L4 (the thalamic input layer) lacks effective illusory contour inference encoding, suggesting that inference-related information arises from cortical circuits rather than bottom-up input. Higher visual areas, such as LM, also demonstrated pronounced illusory contour inference activity.

2. Functional Separation of Different Neuronal Groups:

   • IC-encoders: Specialize in responding to the global illusory contour, receive strong feedback from upstream areas (such as LM), and can recapitulate pattern completion locally within the cortical network.
   • Segment Responders: React to individual inducing segments, facilitate upward transmission of bottom-up input to higher visual areas, but have limited impact on pattern completion for local inference.

III. Causal Regulation of Neuronal Groups and Pattern Completion Mechanism

1. Holographic Optogenetic Manipulation: Using 2p holographic optogenetics, the authors targeted defined IC-encoders or segment responders for selective activation in the absence of visual input, while simultaneously monitoring the activity changes in other neurons within the cortex.

2. Innovative “All-optical Read/Write” Workflow: The experiment was conducted in three steps—first, identifying neuronal functional types (IC-encoders/segment responders); second, online computation to select optical stimulation targets; and third, stimulating the targeted groups and monitoring the global cortical response.

3. Validation of Pattern Completion Effects:

   • IC-encoder Activation Experiment: Upon activation of IC-encoders, the local V1L2/3 cortex network autonomously “fills in” the illusory contour template. Even without actual visual input, other non-activated IC-responsive neurons exhibit activity patterns highly similar to those evoked by real illusory contour stimuli, indicating that local circuits reinforce inference information via the pattern completion mechanism.
   • Segment Responder Activation Experiment: Primarily drives downward signals arriving at higher visual areas but cannot efficiently generate illusory contour pattern creation within V1L2/3.
   • LC-encoder Control Experiment: In contrast, activation of LC-encoders is much less effective than IC-encoders in pattern completion, indicating a unique role for IC-encoding neurons in pattern recreation.

4. Large Field-of-View Mesoscope Multi-Area Verification: Using the self-developed 2p holographic mesoscope, it was further confirmed that segment responders efficiently advance information to higher visual areas, whereas IC-encoders are confined to local V1L2/3 circuits, forming a complete cortical hierarchical circuit division.

IV. Overall Conclusions and Scientific Significance

1. Cortical Mechanism Model of Illusory Contour Inference: For the first time, the research comprehensively elucidates that in the mouse visual system, illusory contour inference is dominated by IC-encoding neurons in V1L2/3, which locally amplify and maintain inference information via recurrent pattern completion mechanisms. Segment responders act as bridges, transmitting bottom-up information to upstream areas and participating in higher-level inference computations.

2. Pattern Completion Theory and Artificial Intelligence Analogy: This study compellingly addresses a key question in neural network science—that pure feedforward linear networks cannot simulate illusory contour perception. Only by introducing cortical loops and recurrent pattern completion can brain-like artificial networks possess real inferential capability. This insight provides a theoretical reference for next-generation AI vision systems.

3. Scientific and Applied Value: The study unveils the cellular circuit basis of neocortical inference function, paving the way for understanding inference mechanisms in human/animal vision, and offering new ideas and targets for disease (such as hallucinations and perceptual disorders) and intelligent algorithm design.

Research Highlights and Innovations

• Systematic Innovation in Experimental Techniques: For the first time, multi-channel high-density electrophysiology, large-field two-photon imaging, and holographic optogenetics are integrated within a single study framework, enabling selective manipulation and causal verification of functional neuron groups.
• Innovative Data Analysis Algorithms: Developed the inference decoding analysis paradigm, enabling precise quantification and hierarchical comparison of the inference capability of neuronal populations.
• Key Findings: Identified V1L2/3 as the earliest cortical region exhibiting illusory contour inference, and distinguished the different circuit functions of IC-encoders and segment responders.
• Novel Pattern Completion Mechanism: Was the first to experimentally demonstrate that neocortical pattern completion is the physical basis of perceptual inference, with inference signals repeatedly activated and persistently reinforced by local circuits.
• Interdisciplinary Theoretical Integration: Organically combines cognitive neuroscience, AI neural networks, and biological theory, providing each field with clear mechanistic models and validation frameworks.

Important Additional Information

Beyond the research methods and findings, the paper details various transgenic mouse lines, surgical procedures, data acquisition (e.g., kilosort2 spike sorting, suite2p motion correction), neuron classification criteria, and extensive literature. The self-developed 2p holographic mesoscope provides a powerful platform for future multi-area and multi-layer causal mechanism validation.

Conclusion: Advancing Research on the Neural Mechanisms of Sensory Inference

Through deep integration of multiple innovative technologies and theoretical analyses, this study systematically reveals the circuit mechanisms underlying neocortical sensory inference, for the first time conclusively establishing the core role of “recurrent pattern completion” in illusory contour perception. It not only provides robust evidence for fundamental cognitive neuroscience, but also opens new pathways for AI, intervention strategies for cognitive disorders, and perception-related diseases. Future explorations based on similar technological platforms and theoretical paradigms hold promise for uncovering more complex perceptual inference behaviors and their neuronal circuit bases, making significant contributions to our understanding of human cognition and the development of brain-inspired intelligence.