Journal of Cognitive Neuroscience
April 15, 2016
Maxine T. Sherman, Ryota Kanai, Anil K. Seth et al.
134 citations
Spontaneous alpha-band neural oscillations in the brain periodically transmit prior expectations to the visual cortex, biasing both objective decisions and subjective confidence before a stimulus appears. In a detection task with scalp EEG, prestimulus occipital alpha phase predicted the weighting of expectations on yes/no decisions and on confidence judgments, independent of attention. These findings suggest that alpha oscillations change the baseline from which evidence accumulation begins, shaping early visual processing and informing how expectations influence perception at the neural level.
May 18, 2023
Arthur Juliani, Adam Safron, Ryota Kanai
9 citations
preprint
Psychedelic therapy shows promise for treating various mental disorders. The REBUS model proposes that psychedelics help by relaxing overly rigid, maladaptive beliefs. The CANAL model extends this by suggesting that canalization—the development of excessively rigid belief structures—may underlie psychopathology. This paper refines the CANAL model by drawing on learning theory from deep neural networks, distinguishing two separate optimization landscapes for belief representation. Each landscape can develop pathologies from either too much or too little canalization, indicating a non-linear relationship with psychopathology. The refined model generates novel predictions about which psychopathologies might respond to psychedelic therapy and which forms of therapy may benefit specific individuals.
Neuroscience of Consciousness
January 1, 2024
Arthur Juliani, Adam Safron, Ryota Kanai
8 citations
Psychedelic therapy shows promise for treating mental disorders, and the "RElaxed Beliefs Under pSychedelics" (REBUS) model explains this by suggesting psychedelics loosen maladaptive high-level beliefs. The newer "CANAL" model proposes that overly rigid belief landscapes (canalization) contribute to psychopathology. This work uses deep neural network learning theory to refine the CANAL model, distinguishing two separate optimization landscapes for belief representation in the brain. Each can develop unique pathologies from either too much or too little canalization, indicating that canalization's link to psychopathology is not simply linear. The refined model makes novel predictions about which aspects of psychopathology psychedelic therapy may treat and which therapy forms might benefit a given individual.
arXiv Preprint Archive
June 13, 2026
Ryota Kanai, Shuqin Ma
A common objection to artificial consciousness—that a simulated brain is no more conscious than simulated water is wet—is addressed from the perspective of Intrinsic Computational Functionalism. Consciousness depends not on external descriptions but on computational structures a system physically realizes through its own causal-dynamical organization. Previous work defined functional states by input-output roles under a fixed interface, but this is incomplete because it makes lookup tables and unfolded systems canonically equivalent.
Trends in Cognitive Sciences
June 1, 2026
Patrick Butlin, Robert Long, Tim Bayne et al.
A method for assessing whether AI systems might be conscious is presented, drawing on existing neuroscientific theories of consciousness. The approach involves deriving indicators from such theories to inform beliefs about AI consciousness. This method can make progress because computational functionalist theories, which are influential, have empirically testable implications for AI. The work does not claim that any current AI is conscious but outlines a rigorous framework for future assessment.
arXiv.org
May 9, 2026
Ryota Kanai, Shu Ma
Canonical functionalism reframes the debate about consciousness by identifying consciousness-relevant functional organization with a system's minimal state-transition structure derived from counterfactual roles, rather than arbitrary input-output mappings or semantic labels. This approach avoids observer-relative interpretations that often plague computational functionalism. The framework does not claim to identify which systems are conscious or that functional organization suffices for consciousness; instead, it specifies the canonical object over which functionalist theories should be formulated. It reframes objections about lookup tables and simulations by requiring functionalists to specify whether the relevant canonical structure is preserved.
Neuroscience of Consciousness
January 1, 2025
Ryota Kanai, Ryota Takatsuki, Ippei Fujisawa
A refined computational interpretation of meta-representations in higher-order theories (HOT) of consciousness is proposed, focusing on process-level representations rather than mere transformations of first-order states. Meta-representations are argued to represent the computational processes that generate first-order representations, building on the Radical Plasticity Thesis. As a proof-of-concept, "meta-networks" were constructed using autoencoders of first-order neural networks within deep learning architectures, where latent spaces embedding first-order networks correspond to meta-representations. Applied to neural networks trained on visual and auditory datasets, these meta-representations successfully captured qualitative aspects by separating visual and auditory networks in the meta-representation space. This formulation offers an empirically testable hypothesis that brain regions may represent processes transforming one representation into another, potentially underlying the ability to describe qualia.
Frontiers in Computational Neuroscience
January 1, 2024
Rousslan Fernand Julien Dossa, Kai Arulkumaran, Arthur Juliani et al.
An embodied agent with a structure based on global workspace theory, trained on realistic audiovisual inputs to navigate 3D environments, performs better and more robustly at smaller working memory sizes compared to a standard recurrent architecture. Task complexity and regularization are essential for feature learning and the development of meaningful attentional patterns within the workspace.
Neuroscience of Consciousness
January 1, 2024
Ryota Kanai, Ippei Fujisawa
The paper introduces 'Universality' as a desirable property for theories of consciousness, borrowed from physics, where fundamental laws apply consistently everywhere. Universality requires that a theory can determine whether any fully described dynamical system is conscious or non-conscious, based on intrinsic properties rather than external interpretation. Most current theories lack this property, as they focus on neural correlates of consciousness in brain-centric systems. The authors argue that functionalist theories could become universal by specifying mathematical formulations of their concepts. While neurobiological and functionalist theories remain useful, a universal theory is needed to fully explain why certain systems possess consciousness.
arXiv Preprint Archive
August 17, 2023
Patrick Butlin, Robert Long, Eric Elmoznino et al.
No current AI systems are conscious, but there are no obvious technical barriers to building ones that might be, according to an analysis grounded in neuroscientific theories of consciousness. The report surveys prominent theories—recurrent processing, global workspace, higher-order, predictive processing, and attention schema—and derives computational indicator properties from them. Applying these indicators to recent AI systems yields no evidence of consciousness, but the authors argue that future systems could potentially implement the necessary properties.
arXiv Preprint Archive
March 24, 2022
Arthur Juliani, Kai Arulkumaran, Shuntaro Sasai et al.
The authors examine three contemporary theories of conscious function—Global Workspace Theory, Information Generation Theory, and Attention Schema Theory—and find that each relates conscious function to some aspect of domain-general intelligence in humans. They then observe that state-of-the-art deep learning methods have begun incorporating key aspects of these theories, though they remain far from demonstrating general intelligence. Using mental time travel in humans as a motivating example, the authors propose combining insights from all three theories into a single unified model. Such artificial agents would possess greater general intelligence and align more closely with current understanding of consciousness's functional role, making this a promising near-term AI research goal.
Trends in Neurosciences
September 1, 2021
Rufin Vanrullen, Ryota Kanai
Deep learning has brought artificial intelligence close to human performance on many tasks, but new brain-inspired architectures are still needed. The Global Workspace Theory describes a large-scale system that integrates and distributes information among specialized modules to support higher-level cognition and awareness. The authors argue it is now time to implement this theory explicitly with deep-learning techniques. They propose a roadmap using unsupervised neural translation between multiple latent spaces—neural networks trained on different tasks or sensory modalities—to create a single, amodal Global Latent Workspace (GLW). The potential functional advantages of GLW and its neuroscientific implications are reviewed.
arXiv Preprint Archive
September 28, 2019
Acer Y. C. Chang, Martin Biehl, Yen Yu et al.
Conscious experience corresponds to information encoded in coarse-grained neural states, such as the firing patterns of neuronal populations, rather than in the noisy activity of individual neurons or in macro-level interactions like interpersonal communication. The authors introduce Information Closure Theory of Consciousness (ICT), which hypothesizes that conscious processes form non-trivial informational closure (NTIC) with respect to the environment at certain coarse-grained levels. This closure confines conscious experience to those levels. ICT provides quantitative definitions of conscious content and conscious level, offering explanations and predictions for various consciousness phenomena and reconciling issues in existing theories.
Neuroscience of Consciousness
January 1, 2019
Ryota Kanai, Acer Y. C. Chang, Yen Yu et al.
Consciousness may have evolved because it enables organisms to internally generate representations of events not tied to current sensory input, using generative models built through sensory-motor interactions. This capacity for information generation supports intention, imagination, planning, short-term memory, attention, curiosity, and creativity, all of which contribute to non-reflexive, flexible behavior. The hypothesis aligns with predictive coding, where top-down predictions correspond to information generation, and with empirical evidence that recurrent feedback activations are linked to consciousness while feedforward processing alone occurs without conscious experience. Thus, consciousness provides a biological advantage by allowing internal simulations that endow organisms with intelligent, adaptive behavior.
Entropy (Basel, Switzerland)
March 6, 2018
Jun Kitazono, Ryota Kanai, Masafumi Oizumi
Integrated Information Theory (IIT) links the amount of integrated information (Φ) in the brain to the level of consciousness, proposing that Φ should be measured across the partition of a system where information loss from partitioning is minimized—the Minimum Information Partition (MIP). Exhaustively searching for the MIP is computationally infeasible for large systems. Previous work showed that if a measure of Φ is submodular, an optimization algorithm can find the MIP in polynomial time, but later versions of Φ are not submodular. This study empirically tested the algorithm on non-submodular Φ measures using simulated and real neural data, finding it identifies the MIP nearly perfectly, enabling practical Φ measurement in large systems.
arXiv Preprint Archive
December 19, 2017
Jun Kitazono, Ryota Kanai, Masafumi Oizumi
Integrated Information Theory (IIT) links consciousness to the amount of integrated information (Φ) in the brain, measured across a system's Minimum Information Partition (MIP). Finding the MIP is computationally expensive for large systems. Previous work showed that if Φ satisfies submodularity, an optimization algorithm can find the MIP in polynomial time, but later versions of Φ are not submodular. This study tests the algorithm on non-submodular Φ measures using simulated and real neural data. The algorithm identifies the MIP with near-perfect accuracy even for non-submodular measures, enabling practical measurement of Φ in large systems.
arXiv Preprint Archive
January 18, 2017
Satohiro Tajima, Ryota Kanai
A new framework proposes that the topological dimensionality of shared attractor dynamics can serve as an indicator of integrated information in continuous dynamical systems, addressing practical and theoretical problems in testing integrated information theory (IIT) of consciousness with neuronal signals. The approach uses delay embedding to reconstruct the effects of unobserved nodes on attractor dynamics, allowing identification of the embedded attractor's dimensionality from partial observations. This topological measure is invariant to general coordinate transformations, extending IIT to continuous systems and relaxing conditions needed for evaluating integrated information in real neural data.
Neuroscience of Consciousness
January 1, 2017
Satohiro Tajima, Ryota Kanai
The integrated information theory (IIT) of consciousness proposes that consciousness corresponds to integrated information within neuronal dynamics, but testing this theory empirically is difficult because it requires observing all elements of a neural system simultaneously. This paper suggests that the topological dimensionality of shared attractor dynamics can serve as an indicator of integrated information in continuous attractor dynamics. Using delay embedding, effects of unobserved nodes can be reconstructed from partial observations, allowing identification of the embedded attractor's dimensionality. The topological dimensionality is invariant to coordinate transformations and thus represents a critical property of integrated information. This approach extends IIT to continuous dynamical systems and relaxes the conditions needed to evaluate integrated information in real neural systems, offering a framework for testing the theory with experimental data.
Neuroimage
October 15, 2014
Kristian Sandberg, Gareth Robert Barnes, Bahador Bahrami et al.
During binocular rivalry, when each eye sees a different image, perception alternates between them. Presenting the images intermittently with gaps longer than about 400 milliseconds slows or stops these alternations, a phenomenon called stabilization. Using magnetoencephalography (MEG) in healthy humans, the study found that perceptual content correlates with modulation of stimulus-specific activity in occipital/temporal brain areas 150–270 milliseconds after stimulus onset, possibly reflecting suppression of the non-perceived image. Stability builds gradually across at least ten trials and involves parietal activity.
Ryota Kanai, Ippei Fujisawa, Shinya Tamai et al.
preprint
Consciousness may have evolved as a platform for general intelligence—the ability to apply knowledge from past experiences to solve novel problems. The paper defines general intelligence and outlines three approaches to building AI systems that achieve it: simulation, combination, and generation. These correspond to proposed functions of consciousness from the information generation theory, global workspace theory, and a higher-order theory where qualia are meta-representations. The authors argue that consciousness integrates specialized generative models into a flexible complex, and that qualia allow an agent to choose which models to apply to new problems. These functions could be implemented as artificial consciousness, enabling systems to generate policies for novel problems with minimal trial and error.