Liquid neural networks for adaptive latent brain-state modeling: A perspective

Abstract

Understanding brain function requires models that characterize neural activity as evolving latent brain states rather than as isolated signal segments. Existing artificial intelligence methods have advanced neural decoding, disease classification, functional connectivity analysis, and brain–computer interfaces. However, brain-state prediction, temporal organization, state transitions, dynamical stability, and intervention responses are often studied as separate problems, limiting a unified understanding of how brain states are represented, organized, transformed, disrupted, and regulated. In this perspective, we examine liquid neural networks (LNNs) as a computational framework for adaptive latent brain-state modeling. We do not argue that LNNs are the only approach to continuous-time modeling. Instead, we focus on how their continuous dynamics and input- and state-dependent temporal adaptation can help link latent trajectories, temporal-scale regulation, state stability, and perturbation responses. We position LNNs in relation to spiking neural networks, recurrent neural networks, transformers, neural ordinary differential equations, and state-space models, and discuss their relevance to multimodal neural signal fusion, cortical hierarchical dynamics, brain-state transitions, dynamic disease prediction, and closed-loop neuromodulation. Overall, LNNs show promise for adaptive latent brain-state modeling, but their scientific and translational value depends on rigorous benchmarking, biological validation, computational feasibility, generalization, and real-world safety.