Abstract
Work on perceptual decision-making has used observations of time-varying weighting of sensory stimuli to make inferences about static decision mechanisms in the brain (e.g., the efficiency of integration and/or the form of decision bounds). However, recent work has also shown temporal weighting behavior during motion discrimination is flexibly adaptable to stimulus statistics (Levi et al., 2018). Here, we manipulate temporal weighting strategy to probe the flexibility of the time course of the sensory-driven and choice-correlated responses in the macaque middle temporal area (MT). In doing so, we test how various forms of signal and noise may (or may not) propagate through the visual system to impact downstream processing and behavior. We performed extracellular recordings in MT using linear electrode arrays while a macaque performed a motion discrimination task. The task was completed under three temporal stimulus conditions: (i) Flat, where motion strength had equal statistical expectation over time; (ii) Late, where motion expectation was higher during later stimulus pulses; and (iii) Early, where motion expectation was higher during early stimulus pulses. Psychophysical reverse correlation confirmed that the observer could change their temporal weighting strategy to match stimulus statistics. We could then evaluate changes in the time course of sensory and choice-correlated responses in MT. If temporal weighting strategy is implemented by decision mechanisms after feedforward sensory encoding, the sensory-driven response should not change. However, if choice probability indicates the readout of MT spikes, temporal weighting should be evident in its time course. We characterized these response components using a generalized linear model (GLM), which allowed for statistical dissection of sensory gain from choice-related activity. Across the three conditions, the time course of choice-related activity did not map straightforwardly onto the psychophysical weighting strategy. We discuss the implications for revisions to simple feedforward models of reading out MT.