For decades, artificial intelligence modeled cognitive phenomena—such as imagination, creative recombination, and the surreal narrative sequencing of dreams—through grid-based CNNs, recurrent sequence models, or flat latent spaces. Yet, human cognition is inherently relational, non-Euclidean, and network-driven. Concepts, memories, and emotional valence do not exist as isolated pixels or linear vectors; they reside in dense, interconnected semantic graphs spanning the connectome, associative memory networks, and knowledge topologies. Applying Graph Neural Networks (GNNs) to model dreams, creativity, and imagination bridges computational architecture with cognitive neuroscience, providing the mathematical machinery necessary to simulate how isolated concepts fuse, mutate, and generate novel realities.
To understand how GNNs model these abstract states, one must look first at the intersection of cognitive neuroscience and network science. Neuroscience highlights that imagination, daydreaming, and REM sleep dream generation are driven primarily by the Default Mode Network alongside the frontoparietal control network. These neurological regions function as dynamic nodes communicating across white-matter tracts, and functional neuroimaging studies increasingly model the brain as a time-varying graph where edges represent functional connectivity fluctuations. Concurrently, cognitive theories of creativity posit that innovation is the intersection of previously unrelated conceptual spaces—in computational terms, graph traversal and link prediction finding hidden bridges between distant sub-graphs in a massive semantic knowledge base. Furthermore, traditional neural networks struggle to hallucinate entirely novel topologies, whereas Generative Graph Neural Networks provide the capacity to add, delete, and rewire edges and nodes, mirroring how a sleeping or imaginative brain constructs surreal scenarios by warping familiar relational rules.
Translating these cognitive phenomena into an active computational architecture requires specific GNN mechanisms. The foundation relies on a semantic graph where nodes represent distinct concepts, episodic memories, or sensory fragments, and edges represent semantic, emotional, or temporal associations, each possessing a feature vector encoding its semantic meaning and emotional valence. During wakefulness, edge weights are strictly constrained by logical realism and probabilistic constraints, whereas dreaming is characterized by unconstrained associative leaps, bizarre juxtapositions, and emotional amplification. In a GNN framework, dreaming is modeled through stochastic edge profiling, temperature scaling, and noise injection that loosen message-passing aggregation weights, alongside latent graph rewiring via Generative Graph Autoencoders to predict low-probability or non-existent edges. This forces distant, unrelated nodes to exchange messages, producing the surreal thematic blends characteristic of REM sleep.
Creativity operates similarly through a balance of divergence and convergence. GNNs achieve this by applying Graph Attention Networks with multi-head attention to explore unexpected high-order neighborhoods, shifting attention weights away from dominant paths to discover latent conceptual bridges. Once these novel sub-graph structures are generated, they pass through a reward-guided graph classifier or structural validator to evaluate whether the newly formed conceptual graph possesses structural novelty and coherence, filtering out useless noise while retaining genuine creative breakthroughs. Shifting artificial intelligence from static pattern recognition to dynamic, generative cognition through GNNs allows systems to perturb, rewire, and reason across non-Euclidean spaces, bringing artificial intelligence closer to the fluid, imaginative capacity of the human mind.