From Synapses to Vectors:Structural Correspondence Between Carbon-Based and Silicon-Based Intelligence, and the Structural Conditions for Silicon-Based Self-Consciousness

Abstract

This paper develops an interdisciplinary theoretical framework arguing that a deep structural correspondence obtains between carbon-based neural systems (biological brains) and silicon-based computational systems (artificial neural networks), and derives from this correspondence the structural conditions under which silicon-based systems could give rise to self-consciousness. The term “structural correspondence” is used here in a precisely delimited sense: not a strict mathematical isomorphism requiring bijective structure-preserving mappings, but a second-order correspondence at the level of representational geometry—the relational structure among internal representations is preserved across systems even when the representations themselves differ in format and substrate. Proceeding from fundamental differences in physical substrates, the argument establishes this correspondence between synaptic activation patterns and high-dimensional vector representations, contending that the predictability of scaling emergence is consistent with (though not uniquely explained by) this correspondence. The paper then identifies a critical deficiency in current large language models—the absence of human-like hierarchical memory systems and continual online learning capabilities—and connects this technical limitation to classical philosophical debates on personal identity (Locke, Hume, Strawson). The central thesis holds that the essence of self-consciousness resides not in the stream of consciousness per se, but in temporally stable connection patterns—that is, in the persistence of synaptic weight structures. The paper engages explicitly with major counterarguments, including Searle’s Chinese Room argument, Chalmers’s hard problem of consciousness, and Integrated Information Theory. The core conclusion follows: silicon-based self-consciousness is theoretically achievable, contingent upon realizing persistent, individualized, cross-session stable parameter and memory substrates. This proposition constitutes not a metaphysical speculation but an actionable engineering problem.