hide
Free keywords:
-
Abstract:
To test the predictive power of brain activity during encoding of a natural scene for subsequent
recognition.
Evoked magnetic fields (EMF) were recorded while subjects performed a delayed matchto-
sample task with briefly presented and masked natural images. On each trial a digitized
image was presented for 37ms on a projection screen and immediately followed by
a pattern mask (1 sec). In the subsequent query phase, the subjects first had to judge
whether they would be able to recognize the image. Then the target and three distractor
images were shown, and the subjects had to indicate the target image (4AFC). During the
first 600 ms of presentation EMF’s were recorded with a CTF 151 channels whole cortex
system. We then used the MEG activity from individual trials to predict the subject’s
behavioural response on that trial. To circumvent the effects of guessing, we used only
those trials in which the subjects confidence judgement and recognition performance
agreed. We tested two classifiers, the partial correlation (PC) of a trial with the mean vectors
of correct and false trials and Support Vector (SV) classification which seeks a separating
hyperplane by maximizing the distance to the nearest samples of each class.
In 76.7 of the trials the subjects confidence judgement and response agreed. In 78 of
these trials the subjects gave a correct response. With PC on average 75.3 of the correct
and 76.3 of the false trials were correctly classified. But for only one subject best classification
was obtained with PC. Classification by support vector machines were typically
about 10-15 better than with the PC classifier. Average performance with the best support
vector classifier was about 90.7 for the correct and about 92.8 for the false trials.
It is possible, with about 90 accuracy, to predict in single trials subjects’ subsequent
recognition performance from the early information in the evoked magnetic fields
recorded while subjects were viewing the stimulus.
