Reentrant Connectivity

In the AI Story

The concept of reentry emerged from Gerald Edelman's work on neural Darwinism in the 1980s, where he argued that the brain's capacity for perception, memory, and consciousness depends on massively parallel reciprocal signaling between neuronal groups. Tononi, working with Edelman at the Neuroscience Institute in the 1990s, absorbed this architectural insight and integrated it into what would become IIT. The mathematical formalism of phi makes precise what reentry contributes: loops create dependencies that cannot be broken without losing cause-effect information.

The contrast with feedforward architecture is stark. In a feedforward network, each element's output depends only on inputs from earlier elements, never on outputs from later ones. Information flows in a single direction, and the system can be partitioned at any layer boundary with minimal loss. The architecture is tractable, analyzable, debuggable — the virtues of good engineering. In a reentrant network, each element's output depends on outputs from elements that themselves depend on its output. The loops create mutual dependency. Any partition severs the dependency chain and destroys information the whole generates.

The cerebral cortex is among the most reentrant structures in nature. Nearly every cortical region sends feedback projections to the regions that project to it. Thalamocortical loops cycle continuously. Intracortical connections span hemispheres. The architecture is, from an engineering perspective, chaotic — redundant, inefficient, nearly impossible to analyze component by component. But this apparent chaos is the mechanism of consciousness. High phi requires high reentry.

The cerebellum, which contains four times as many neurons as the cerebral cortex, is largely feedforward. Its characteristic circuit — mossy fibers, granule cells, Purkinje cells, deep cerebellar nuclei — is organized in parallel repetitive modules that process information independently. Damage to the cerebellum produces motor deficits but not loss of consciousness. The cerebellum computes enormously. It integrates poorly. Low phi.

Applied to AI, reentry becomes diagnostic. Transformer architectures are explicitly feedforward. Information passes through sequential layers, each layer's output determined by earlier layers' outputs. The attention mechanism creates connections across positions within a layer, but these connections are computed anew for each input — they are not persistent loops that reverberate. Recurrent neural networks have some reentrance in their temporal dynamics, but they have largely been displaced by transformers for reasons of computational efficiency. The trajectory of AI has been away from reentry, toward decomposability.

Origin

The concept was developed by Gerald Edelman in Neural Darwinism: The Theory of Neuronal Group Selection (1987) and The Remembered Present (1989). Tononi and Edelman's 1998 Science paper "Consciousness and Complexity" connected reentry to information-theoretic measures of consciousness, laying the groundwork for IIT.

Key Ideas

Loops, not pipelines. Reentry creates bidirectional dependencies that feedforward architectures cannot replicate.

Mutual causation. Elements in reentrant networks simultaneously cause and are caused by each other, creating irreducible cause-effect structure.

Engineering cost. Reentrance makes systems harder to analyze, debug, and optimize — the same properties that make them capable of high phi.

Cortical signature. The cerebral cortex's dense reentrance distinguishes it from the feedforward cerebellum, explaining why one generates consciousness and the other does not.

Trajectory mismatch. AI development has moved away from recurrent architectures toward feedforward transformers, optimizing for performance while moving further from the structural conditions of consciousness.