Computational theories of attention in decision making

For many years, research on attention has been dominated by theories based on the assumption that attention is limited in capacity. These include the limited-capacity channel theories of Welford and Broadbent and capacity or resource theories by Moray, Posner, and Kahneman. This talk challenges those theories and their many descendants by asking why capacity is limited and what role capacity plays in the computations required to perform attention tasks. There are few satisfactory answers in limited-capacity theories of attention. I show that the effects of load on performance, which are commonly interpreted as evidence for limited capacity, can be produced by models that assume unlimited, limited, and fixed capacity. I argue that attention is better construed as selection of information that we need to achieve our goals. Following current research on computational models of attention in associative learning, categorization, perceptual learning, cognitive development, neuroscience, and artificial intelligence, I propose that attention is a process of choice, in which selection implemented as multiplicative gain control and processing is constrained by normalization. This perspective focuses on interactions between representations and decision processes applied to them, explaining many attentional phenomena without assuming attention has limited capacity.

This is an in-person presentation on July 20, 2026 (09:00 ~ 09:20 EDT).

Selective attention is typically treated as a dedicated cognitive mechanism that modulates representations to favor task-relevant information. We challenge this view, presenting evidence from two very different modeling frameworks. We find that attentional effects can arise as a natural consequence of adaptive processing rather than requiring purpose-built attentional machinery. In the first line of work, we analyze monkey spiking data from a task requiring trial-by-trial switching between color and motion judgments. Representations across all recorded cortical areas, including V4, MT, PFC, FEF, LIP, and IT, stretched along the task-relevant dimension, with spike timing carrying critical information. An LSTM deep network trained on identical input sequences and rewards, without any explicit attentional mechanism, displayed the same qualitative stretching as a consequence of minimizing prediction error. In this case, selective attention emerges from error minimization. In the second line of work, we develop the Sampling Emergent Attention (SEA) model, which re-conceptualizes attention through a Bayesian lens as the expected utility of sampling particular information sources. Attentional effects emerge from a cost-sensitive information-gathering process, including flexible, stimulus-specific allocation patterns that mirror human eye-tracking data and go beyond what traditional fixed-weight attention models predict. In this case, selective attention emerges from utility maximization. Despite their fundamental differences, both frameworks converge on the same conclusion: attentional phenomena do not imply attentional mechanisms. Whether optimizing prediction error or expected utility, adaptive systems naturally develop the representational and behavioral signatures of attention, which invites a reappraisal of what attention fundamentally is. References http://dx.doi.org/10.1038/s41467-025-65231-y https://doi.org/10.1037/rev0000287

This is an in-person presentation on July 20, 2026 (09:20 ~ 09:40 EDT).

Attention has been criticised for its lack of theoretical cohesion. We show that a single computational framework - the adaptive representation model – can unify several competing conceptualizations. Specifically, we model attention in a category learning task where the underlying category structure changes. In our model, cognitive agents pay attention to features to weight them to make category decisions. These attention weights are dynamic and updated using gradient descent to maximize performance on every trial. Features that are attended to are stored in memory and used to inform the attentional update. Thus, the same attentional mechanism is used to (i) weight dimensions based on perceived importance that influences the (ii) encoding of relevant information (iii) explain adaptive behavior such as rapid phase transformations during learning and (iv) explain maladaptive behavior like catastrophic transfer and learning traps. We construct a switchboard framework to turn these different mechanisms on and off and show their contribution to explaining behavior. We test our computational account using empirical cross-species experiments involving humans, rats and pigeons. We find evidence for selective attention for humans but not rats and pigeons. Overall, we find that despite its seeming lack of theoretical cohesion, jointly conceptualizing attention is critical to developing a coherent account of learning, memory and decision making.

This is an in-person presentation on July 20, 2026 (09:40 ~ 10:00 EDT).

Episodic memory can recall detailed event representations from sparse cues, yet information learned during an event can also be generalized to novel situations. To explain this flexibility, many accounts posit multiple memory systems or the storage of separate traces for specific and generalized content. Here, we argue that single-system, single-trace architectures may be substantially more flexible than typically assumed. We show that global matching models of episodic memory and exemplar models of classification can be viewed as static approximations of a recurrent network architecture that we call ATHENA. Simulations demonstrate that plausible modulations of network components alter the effective level of competition between traces during retrieval. In traditional cognitive models, this corresponds to dynamically varying the width of the similarity kernel. Allowing this variation enables a single-system, single-trace model to retrieve individual event memories (i.e., exemplars), averages of related memories (i.e., prototypes or clusters), and global averages across all memories (i.e., base rates of features) using the same memory traces and retrieval cue. We conclude that, once static-time assumptions are relaxed, episodic memory systems can naturally support both precise recall and flexible generalization.

This is an in-person presentation on July 20, 2026 (10:00 ~ 10:20 EDT).