Human perception and experience of time is strongly affected by environmental context. When paying close attention to time, time experience seems to expand; when distracted from time, experience of time seems to contract. Contrasts in experiences like these are common enough to be exemplified in sayings like "time flies when you're having fun". Similarly, experience of time depends on the content of perceptual experience -more rapidly changing or complex perceptual scenes seem longer in duration than less dynamic ones. The complexity of interactions among stimulation, attention, and memory that characterise time experience is likely the reason that a single overarching theory of time perception has been difficult to achieve. In the present study we propose a framework that reconciles these interactions within a single model, built using the principles of the predictive processing approach to perception. We designed a neural hierarchical Bayesian system, functionally similar to human perceptual processing, making use of hierarchical predictive coding, short-term plasticity, spatio-temporal attention, and episodic memory formation and recall. A large-scale experiment with ∼ 13, 000 human participants investigated the effects of memory, cognitive load, and stimulus content on duration reports of natural scenes up to ∼ 1 minute long. Model-based estimates matched human reports, replicating key qualitative biases including differences by cognitive load, scene type, and judgement (prospective or retrospective). Our approach provides an end-to-end model of duration perception from natural stimulus processing to estimation and from current experience to recalling the past, providing a new understanding of this central aspect of human experience.
Predictive processing neural architecture for perceptionAccording to (Bayesian) predictive processing theories, the brain is constantly updating its internal representations of the state of the world with new sensory information x 0 t (at time t) through the process of hierarchical probabilistic inference 14,17,19,20 . The resulting 'top-down' generative model is used to predict new states and simulate the outcome of prospective actions 21 . Comparing predicted against actual sensory signals gives rise to prediction errors ξ n t which, in most instantiations of predictive processing, are assumed to flow in a 'bottom-up' direction from the sensory periphery (n = 0), towards higher hierarchical levels n ∈ {1, 2, ..}. Numerous connections with neurophysiology have been made over the years, including proposals viewing predictive processing as the fundamental function of canonical microcircuits in the cortex 22 , or as the result of the interplay between competing oscillatory bands 23, 24 , as well as providing theoretical explanations for a wide range of cognitive and perceptual effects such as hallucinations 25 , autism 26 and schizophrenia 27 . In addition, this framework is uniquely qualified for studying the relation between perceptual change and human perception of time 4 , as i...