Intrinsic neural timescales (INT) provide a key temporal architecture for cognition, underpinning hierarchical information processing in the brain. Altered INTs are a hallmark of neuropsychiatric disorders, particularly schizophrenia, where they contribute to characteristic deficits in cognitive flexibility and predictive coding.
This thesis employs the Hierarchical Gaussian Filter (HGF), a Bayesian learning framework, to mechanistically investigate how specific parameter variations shape belief updating and govern the emergence of hierarchical timescales. Simulations of two- and three-level perceptual models establish the HGF’s inherent capacity
to generate a hierarchy of timescales with increasing temporal value at higher levels. We then calibrate these parameters using clinical data to simulate disease-specific learning architectures. Compared to healthy controls, schizophrenia-like agents exhibit shorter INTs, heightened sensitivity to uncertainty and elevated initial estimates of environmental log-volatility, in line with their characteristic mechanism of impaired flexible decision-making. These computational phenotypes align with empirical findings of abnormal temporal integration and maladaptive belief updating, providing a mechanistic link between altered intrinsic timescales and deficits in predictive coding for schizophrenia patients.