Adaptive allocation of attentional gain

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant): Human sensory systems cannot simultaneously parse and reconstruct all available inputs into meaningful perceptual representations. Classic accounts address this processing limit by invoking a selection mechanism that preferentially encodes only the most salient and behaviorally relevant stimuli in the environment. This mechanism is typically referred to as selective attention, and empirical work has traditionally focused on understanding why and how relevant stimuli dominate perceptual awareness. However, other extra-retinal factors can also impact the efficiency of information processing, such as estimates of the prior probability of a particular stimulus (or configuration of stimuli) based on past experience in a particular context (i.e. expectation). Despite many demonstrations that expectation can profoundly influence a variety of perceptual phenomena ranging from low-level grouping to high-level object recognition, empirical and theoretical studies almost always conflate expectation and selective attention. This confusion persists even though these factors are logically dissociable: the probability that a stimulus will appear in a given context may have little or nothing to do with behavioral relevance. The conflation of these extra-retinal factors may seem inconsequential, as both might naively be expected to influence neural activity and behavior in a similar way. However, recent theories of cortical information processing - such as predictive coding - hold that stable perceptual representations emerge from the dynamic interplay between internal probability estimates about the state of the world (i.e. expectations) and the content and quality of incoming sensory information (which is shaped by task-relevance, or attention). Here, we adopt a Bayesian framework that casts perceptual inference as the product of prior beliefs and likelihoods (i.e. sensory evidence). We will use this framework to formulate and test the hypothesis that expectation operates on priors to modulate pre-stimulus responses in visual cortex and to bias the `read-out' of neural codes during decision-making, whereas attention directly impacts likelihood functions by shaping stimulus-evoked neural responses on the basis of task relevance. Our approach will combine psychophysics, quantitative models of perceptual and cognitive processes, and novel EEG and fMRI analysis methods that can determine how priors and likelihoods combine to shape the quality of feature-selective perceptual representations. Collectively, this work will provide key insights into how different extra-retinal biasing factors interact to shape perception, and will more broadly test generative models of cortical information processing that characterize perception as a problem of optimal statistical inference. In turn, this knowledge should improve our ability to isolate specific aspects of selective information processing that can sometimes go awry, thereby enabling more targeted diagnoses and interventions in clinical settings.