P-9 Difference in readout, not encoding, of sensory evidence for continuous decisions dependent on consistency with previous categorical judgment

Background: Many perceptual decisions are based on the accumulation
of protracted streams of sensory evidence. Recent behavioral
work has required participants to produce a continuous estimate
of the mean of a stream of visual evidence, combined with a categorical
judgment in the middle of the stream. In such tasks, human subjects
put stronger weight on subsequent evidence samples that are
consistent compared to inconsistent with that intermittent categorical
judgment. The neural basis of this cognitive bias is unknown.
Objective: We asked whether changes in evidence weighting
depending on consistency with previous categorical judgment stems
from (i) encoding of the evidence in visual cortex or (ii) in the readout
of this information for a final continuous estimation report.
Methods: During magnetoencephalography (MEG), 34 healthy
human participants performed an estimation task based on a stream
of visual evidence: the angular positions of checkerboard patches
drawn from a Gaussian distribution (trial varying mean: 1414,
fixed s.d.: 20). The task was to report the sample mean (mean of the
stream of visual stimuli) by moving a cursor on the screen at the end
of the trial. Importantly, after seeing half of the stream, they made a
categorical judgment (the mean is on the left vs. right) about the
visual stimuli or viewed a left vs. right cue (75% validity) about the
mean of the stream. We reconstructed the MEG responses following
sample onsets for areas across the visual cortical hierarchy and used
information-theoretic measures to quantify the relationship
between the local response pattern and evidence sample and final
estimate. We also quantified a statistical signature of sensory readout:
the ‘intersection information’ between evidence sample, neural
response, and estimation.
Results: At the behavioral level, participants gave stronger weight
for samples consistent compared to inconsistent with the previous
intermittent categorical judgment or cue. At the neural level, there
was no difference in the encoding of evidence samples (quantified
as mutual information) between consistent and inconsistent samples
in any visual cortical area. Critically, the intersection information
in early visual cortex (areas V2-V4, similar but weaker in V1)
and downstream areas was larger for evidence samples consistent
than inconsistent with the categorical choice or cue.
Conclusion: The consistency between categorical information (own
judgment or external cue) and subsequent evidence samples does
not change the encoding of the evidence sample in early sensory cortex,
but selectively modulates the readout of these sensory evidence
signals in the downstream processing underlying the behavioral
estimation report. This differential sensory readout could be due to
differences in the routing of the sample information to downstream
circuits involved in integrating the samples into an evolving decision
variable for the estimation, in the updating of that decision variable
by the new sample, or both.