Eivinas Butkus^1,2, Zhuofan Ying^1,2, Nikolaus Kriegeskorte^1,2; ¹Columbia University, ²NSF AI Institute for Artificial and Natural Intelligence

Primate vision is thought to rely on attention to overcome capacity limitations, including severe energetic constraints. Yet how attention enables efficient use of energy—and whether the benefits outweigh the costs of attentional machinery—has never been rigorously demonstrated. Here we provide a mechanistic account of how attention supports energy-efficient vision and flexible energy-accuracy trade-offs. Our model, EAN (“Energy-efficient Attention Network”), implements attentional selection as recurrent top-down multiplicative gain varying over features, space, and time within a convolutional neural network. We optimize a joint objective balancing high accuracy and low energy costs. The energy cost accounts for action potentials and synaptic transmission throughout all model components, including the attention controller. We test EAN on a novel visual-category-search (VCS) task requiring identification of a digit's class (what, 0-9) and location (where) among distractor letters. EAN learns to attend broadly in an initial low-energy pass and then dynamically focuses its energy on task-relevant locations and features of the particular scene at hand. The attention mechanism substantially improves energy efficiency and enables flexible trading of accuracy against energy on a trial-by-trial basis. The model variant with both feature and spatial gain is most efficient and best explains human errors and difficulty judgments. EAN readily generalizes to classical attention tasks and captures canonical single-cell electrophysiological attentional effects on firing rate, Fano factor, and noise correlations. Our work connects a cognitive function (attention), a neural mechanism (gain modulation), and a neurobiological constraint (metabolic costs) in a unified mechanistic account of how recurrence and selection enable flexible, energy-efficient vision.