Sense of Coherence in the Perinatal Period: A Longitudinal Growth Mixture Modeling Analysis.

Seeing a speaker’s lip while listening to his voice is known to enhance language understanding, which exemplifies the computational power of human brain to synthesize disparate sensory modalities into a unified percept through multisensory integration. This process generates emergent perceptual qualities unattainable from any single modality, leading to improve the perceptual accuracy, accelerate decision-making, and underlay a range of psychophysical phenomena. Among varied sensory modalities, audio and visual inputs dominate over 90% of our interactions with the external world, underpinning critical cognitive functions from spatial navigation to attentional selection. Recent advances in neurophysiology have begun to uncover how multisensory neurons in biological brains can integrate audiovisual inputs from synaptic-level plasticity to large-scale cortical interactions. At its core, multisensory neurons integrate stimuli through three key principles: superadditivity (i.e., integration responses surpassing the sum of unisensory inputs), inverse effectiveness (i.e., maximal integration for weak unimodal signals, and minimum for strong ones), and temporal congruency (i.e., temporally aligned cross-modal inputs for integration). Harnessing these multisensory-integration mechanisms could revolutionize artificial intelligence by enabling self-adaptive perceptions for robotics, autonomous vehicles, and smart homes. This has long been a frontier in neuromorphic and electronic engineering, yet facing huge challenges. Current approaches primarily rely on combining multiple sensors with algorithmic frameworks such as weighted averaging, feature concatenation, and principal component analysis, whereas lacking the multiplicative nonlinearity and adaptive reweighting intrinsic to biological neuronal networks. While artificial neural networks approximate the superadditivity via nonlinear activation functions, they struggle to autonomously implement the inverse effectiveness that dynamically reweights synaptic inputs based on real-time reliability of signals. Since biological multisensory neurons exploit the inverse effectiveness through self-tuned thresholds, most algorithmic fusions typically emulate this inverse behavior by hand-engineered rules or complex Bayesian inference layers that impose substantial computational overhead. Moreover, maintaining the temporal congruency amid dynamically-reweighted inputs is non-trivial, as algorithmic synchronizations often conflict with the adaptive processing. These limitations are becoming particularly pronounced in edge hardware, where stringent energy-efficiency requirements are imposed on computational resources. Recent advances in physical computing offer promising potential to implement neural principles directly in hardware through device physics, yet to be fully explored. Here, we demonstrate biomimetic ferroelectric-semiconductor field-effect transistors (FeS-FETs) that natively emulate audiovisual integration at the device level. By exploiting the coupling dynamics of ferroelectric polarization and semiconducting photocurrent in the ferroelectric-semiconducting BiOSe channel, our FeS-FETs simultaneously implement neural integration features of superadditivity, inverse effectiveness, and temporal congruency within a single compact architecture. In particular, an ultrahigh superadditive integration factor exceeding 2800% is achieved alongside the prolonged temporal congruency over 10 s. Furthermore, we integrate the FeS-FET array with memristor-chip-based spiking neural networks (SNNs) to construct a hierarchical multisensory recognition system, which replicates versatile neurobiological behavior, including the sensory synaptic plasticity, population-coded spiking, and Bayesian-optimal fusion. This multisensory system reaches an excellent recognition accuracy of 98.2% for fuzzy objects, outperforming conventional unimodal and algorithmic-fusion approaches. Through adopting the intrinsic multi-physical dynamics of BiOSe FeS-FETs, our work introduces a hardware-adaptive paradigm for edge-compatible multisensory intelligences.