ITERATIVE FEATURE-BASED ATTENTION ENABLES HYPOTHESIS DISCRIMINATION VIA REPRESENTATIONAL ADAPTATION

Perceptual accuracy often improves with processing time, even for static stimuli that provide no additional sensory evidence. While this speed-accuracy trade-off is well documented behaviorally, its representational and computational basis remains unclear: what changes over time when the input itself is unchanged? We propose that additional processing time enables iterative attentional re-representation of sensory input, allowing the same neural architecture to become differentially effective across contexts and tasks without structural changes to synaptic weights or connectivity. In this view, attention supports temporary representational adaptation, altering which sensory distinctions are emphasized to better support hypothesis identification. Using a deep neural network as a model of visual inference, we study this idea in the context of category discrimination, treated here as a concrete instance of a more general problem: differentiating among competing hypotheses about how sensory data should be interpreted or acted upon. Building on prior work on feature-based attention (e.g., Martinez-Trujillo & Treue, 2004, Current Biology 14, 744-751; Lindsay & Miller, 2018, eLife 7, e38105), we introduce a feedback-driven attentional mechanism trained to minimize entropy over competing hypotheses by modulating intermediate feature representations during stimulus processing. This targeted entropy reduction increases mutual information between sensory representations and task-relevant hypotheses, improving their discriminability. Iterative representational refinement yields higher inference accuracy for static inputs, providing a functional and mechanistic account of how processing time improves performance, with implications beyond object recognition for hypothesis-driven perception and decision making.