Abstract Human language processing, characterized by its ability to capture long-range dependencies in sequential inputs, operates under the constraints of limited working memory. In contrast, state-of-the-art transformer models in artificial intelligence rely on access to the fixed context window, which deviates from the dynamic nature of human cognition. Here, we propose a novel approach to reconcile this disparity by integrating a computational model of working memory into the transformer architecture. This biologically-inspired modification constructs a time-local transformer, capable of learning complex dependencies without needing the full input history. Our findings demonstrate that this approach still preserves the capacity of transformers for effective sequence processing. This work is a step towards developing AI models that align more closely with the principles of human brain function, opening new avenues for understanding the neural underpinnings of language and cognition.

Abstract Research in event cognition has shown that how people segment events plays an important role in how they are perceived and remembered. One potential source of information for anticipating upcoming event boundaries is the typical duration of the event. The small amount of prior work that exists suggests that the durations of daily events have normal distributions and that people have accurate prior beliefs about these durations (Griffiths & Tenenbaum, 2006). This research study aims to further explore the temporal distribution of self-reported events from daily life. Forty-eight participants provided information regarding the duration of events from their daily life for 14 days. Event durations for various activity types were characterised using duration modelling with truncated normal, exponential, and gamma models. Predictions based on mean and median values, including ± 5% and ± 10% error margins, were obtained, and the cumulative hazard functions were derived. Exponential and gamma distributions fit most event categories better than the truncated normal distribution. Mean-based predictions of boundaries based on duration showed only 4–5% accuracy, indicating limited usefulness. This aligns with the cumulative hazard results, where most events followed the exponential distribution, suggesting no optimal time for predicting the end of the event. This suggests that when daily events are studied in an ecological context, most of them have little sign of a typical duration. Consequently, it is unlikely that duration estimation plays a large role in the anticipation of event boundaries.

The problem of model selection with a limited number of experimental trials has received considerable attention in cognitive science, where the role of experiments is to discriminate between theories expressed as computational models. Research on this subject has mostly been restricted to optimal experiment design with analytically tractable models. However, cognitive models of increasing complexity with intractable likelihoods are becoming more commonplace. In this paper, we propose BOSMOS, an approach to experimental design that can select between computational models without tractable likelihoods. It does so in a data-efficient manner by sequentially and adaptively generating informative experiments. In contrast to previous approaches, we introduce a novel simulator-based utility objective for design selection and a new approximation of the model likelihood for model selection. In simulated experiments, we demonstrate that the proposed BOSMOS technique can accurately select models in up to two orders of magnitude less time than existing LFI alternatives for three cognitive science tasks: memory retention, sequential signal detection, and risky choice.

The circular diffusion model represents continuous outcome decision making as evidence accumulation by a two-dimensional Wiener process with drift on the interior of a disk, whose radius represents the decision criterion for the task. The hitting point on the circumference of the disk represents the decision outcome and the hitting time represents the decision time. The Girsanov change-of-measure theorem applied to the first-passage time distribution for the Euclidean distance Bessel process yields an explicit expression for the joint distribution of decision outcomes and decision times for the model. A problem with the expression for the joint distribution obtained in this way is that the change-of-measure calculation magnifies numerical noise in the series expression for the Bessel process, which can make the expression unstable at small times when the drift rate or decision criterion is large. We introduce a new method that uses an asymptotic approximation to characterize the Bessel process at short times and the series expression for the large times. The resulting expressions are stable across all parts of the parameter space likely to be of interest in experiments, which greatly simplifies the task of fitting the model to data. The new method applies to the spherical and hyperspherical generalizations of the model and to versions of it in which the drift rates are normally distributed across trials with independent or correlated components.