In theories of consciousness, the question of how awareness emerges from neural activity remains unsettled. Higher-order theories (HOT) propose that conscious experience depends not solely on first-order sensory representations but on meta-representations - representations about those representations. Yet this idea, while conceptually appealing, has long resisted computational definition. The study by Kanai et al. offers a refined, testable formulation: consciousness may depend on representations of processes, not merely on duplicated data streams or confidence estimates about perception.
The team approached the challenge by distinguishing between the states a network holds and the transformations that produce them. In artificial intelligence, deep networks already simulate layered processing similar to the brain's hierarchy, where early layers extract basic features and deeper layers integrate them into coherent patterns. If every layer were simply deemed a "representation of a representation," the notion of meta-representation would lose meaning. Kanai and colleagues therefore proposed a stricter interpretation - meta-representations as structured reflections of the computational procedures that generate first-order representations.
To test this, they built meta-networks - artificial neural systems designed to learn about other networks. Each meta-network received the internal weight patterns of smaller, first-order networks trained on visual and auditory datasets. Instead of analyzing images or sounds directly, the meta-network examined the structure of learning itself, embedding the functional "fingerprint" of each first-order model into a latent space. This architecture allowed the researchers to study how networks representing different sensory modalities diverge at a process level.
The results were striking in their clarity. Using the latent embeddings produced by the meta-network, the system could predict with 100 percent accuracy whether a given first-order network had been trained on visual or auditory data. When visualized through t-SNE dimensionality mapping, the networks clustered into two distinct domains - one for sight, one for sound - without the meta-network ever seeing the original stimuli. In contrast, when classifiers analyzed the raw weight matrices directly, prediction accuracy fell to chance. This demonstrated that the meta-representation space captured qualitative distinctions in how each network processed information.
Beyond modality, the meta-network could also distinguish specific stimulus categories - such as different visual objects or audio events - at levels well above random expectation. These findings indicate that the weight configurations of trained networks encode characteristic process signatures that a higher-order system can detect. In biological terms, the result suggests that the brain might similarly develop representations of its own processing routines - an idea that parallels Cleeremans's Radical Plasticity Thesis, in which consciousness arises from the brain's capacity to learn about its own operations through continual self-modeling.
The researchers emphasize that such meta-representations differ fundamentally from simple confidence estimates or probabilistic "awareness" measures often used in psychology. While a neural network's softmax output may quantify certainty, it lacks the intrinsic "aboutness" required for conscious representation. By contrast, a representation of a process contains relational structure - it encodes how input becomes output, not just the likelihood of correctness. This process-level encoding, the authors argue, provides a computational foothold for describing the qualitative texture of experience, often referred to as qualia.
Kanai et al. further suggest that some brain regions could function as natural meta-networks, learning the transformations occurring between sensory areas rather than the sensory content itself. Such a system could allow the brain to compare, evaluate, and even describe its own processing pathways. This view also resonates with experiments in neuroplasticity: when retinal signals in newborn ferrets were rerouted to the auditory cortex, neurons in that area developed visual-like receptive fields and behaviors, implying that modality may depend more on local wiring and transformation patterns than on anatomical location. The same principle underlies the new model - process defines perception.
The authors note parallels with Integrated Information Theory (IIT), which links consciousness to the intrinsic causal structure of a network. Their approach complements IIT's qualitative perspective by demonstrating that different process architectures naturally separate in representational space, without explicitly computing integrated information. The capacity to predict modality purely from connection patterns supports the notion that qualitative experience corresponds to the organization of transformations within a system.
In cognitive terms, meta-representations provide functional advantages. They create a higher-dimensional map of how subsystems operate, allowing dynamic coordination, error monitoring, and adaptive reconfiguration. For both brains and machines, such self-referential modeling enhances cognitive flexibility: a system aware of its own processing style can select, inhibit, or re-weight internal operations to meet changing demands. This capacity for internal differentiation - the ability to perceive differences within one's own structure - may underlie the subtlety of conscious awareness.
While the study stops short of claiming biological equivalence, its implications extend to the philosophy of mind. If first-order networks correspond to perceptual mechanisms and meta-networks to awareness of those mechanisms, then consciousness may be understood as an emergent informational geometry - a field of relations among processing processes. The authors even venture that qualia themselves might correspond to positions within a "meta-representational space," where similarities between processes define similarities between subjective experiences.
From the perspective of Seven Reflections' Dimensional Systems Architecture (DSA), the study exemplifies a structural recursion between operational layers. DSA models cognition as fields that alternate between expression (first-order) and observation (second-order). Kanai's meta-networks formalize this alternation mathematically: each process becomes both actor and object within a recursive hierarchy. In DSA terms, the shift from representing data to representing transformation corresponds to a transition along the temporal axis - from static structure to dynamic self-mapping. Consciousness, viewed through this lens, arises when a system stabilizes this recursive feedback - when the process perceives its own structure without collapse.
The broader implication is that awareness may not be a product added to computation but an organizational state within it. When systems begin to encode their own transformations, they achieve structural closure: a form of self-consistency that mirrors the architecture of thought. The study thus bridges neuroscience, artificial intelligence, and systems theory, revealing that the path to modeling consciousness may lie not in replicating sensory content but in capturing the laws that govern transformation itself.
