Association between changes in working status and frailty across different types of employment: a longitudinal study in Korea, the world's most rapidly ageing nation.

Animals track their position and orientation as they navigate through uncertain and dynamic environments. To do so, neural populations maintain internal representations of navigational variables—such as head direction (HD) [,]—even in the face of noisy or conflicting sensory cues [–]. Across species, cue uncertainty modulates population-level HD representations, for example broadening an activity bump in flies and altering network gain and latent structure in mice [,]. In general, the brain faces the challenge of integrating uncertain inputs while maintaining coherent internal estimates of external variables and their uncertainties over time [–]. Here, we propose a spiking network that addresses this challenge and employ it to shed light on the HD system’s representation dynamics in the face of uncertainties across species. When faced with uncertain inputs, neural circuits balance two demands: inferring the most likely external signals, potentially while tracking uncertainty [,], and maintaining stable, persistent outputs for computation and behavior []. Different computational perspectives have been used to describe these functions mechanistically. On the one hand, sampling-based probabilistic inference describes how neural variability can be harnessed to explore the posterior distribution over external signals. In sampling, neural variability is not merely noise, but it reflects a specific computation that generates a sequence of possible stimulus values whose statistics approximate the posterior. This has been used extensively to model the early stages of environmental stimulus processing [–], and neural sampling-based mechanisms have accumulated substantial experimental support over the last decade [–]. On the other hand, attractor network models explain how recurrent connectivity can stabilize neural activity patterns, supporting persistent representations of navigation variables such as head direction (HD) [,–]. Biologically, there is substantial evidence that attractor-like mechanisms are at play, with prominent examples in HD systems of rodents [,] and the central complex of the fruit fly [,,]. How can an HD system represent a navigational variable while maintaining and relaying information about the uncertainty of sensory stimuli? In this work, we will consider the possibility that sampling-based inference and attractor dynamics might both be necessary to model higher-order navigation systems. While stochastic fluctuations have long been incorporated into attractor models of decision making [], perceptual rivalry [,], and multistable memory [,], they are usually treated as a perturbation rather than as an explicit mechanism for probabilistic inference. On the other side, prior work has explored stochastic neural dynamics as a mechanism for sampling-based inference [,,,,], but without explicitly incorporating structured continuous attractor manifolds as computational primitives. Here, we set out to construct a network in which: a) stochastic fluctuations play a functional role in implementing sampling-based inference; b) inferred quantities are integrated and attracted to a continuous attractor manifold; c) attractor dynamics are derived from a prior which represents the manifold; and d) input uncertainty is represented and broadcast downstream.