UC San Diego

Visual attention facilitates faster and more accurate decision-making for behaviorally relevant stimuli. To understand the computational mechanism underlying attention, I investigated the impact of attention on population codes in early visual cortex. First, I lay the theoretical foundation of this investigation, whereby the known neuromodulatory activity of attention is examined within the context of information theory to propose a framework termed `Dynamic Neural Coding'. The framework suggests that attention dynamically alters neural codes used to represent basic stimulus features so that behaviorally relevant stimuli are represented and communicated by population responses with higher fidelity than irrelevant stimuli. The framework also suggests that attention and adaptation might jointly mediate neural codes to achieve metabolically efficient sensory information processing. The framework is supported by two experimental studies that use a combination of visual psychophysics, functional magnetic resonance imaging (fMRI), computational modeling, and information-theoretic data analysis, to show how attention modulates population codes to impact information processing. The first experiment reveals that attention increases the quality of sensory representations - measured by an increase in mutual information between population response and stimuli - as early as primary visual cortex. This increase in sensory representation is largely driven by multiplicative scaling of tuning function that forms the population code. According to dynamic neural coding framework, attention- induced improvement in population codes in upstream areas should also improve the transmission of encoded information to downstream areas. The second experiment tested this hypothesis, and found that attention improved the efficacy of communication between two cortical areas - V1 and MT - that interact with during motion processing. Furthermore, a simulation of the inter-cortical interaction between V1 and MT using a computational model reveals that attention-modulations in V1 dictate the fidelity of representations in MT as well as the synchrony between the two areas