Human learning improves machine learning: neural and computational mechanisms of perceptual training.

Full text

Page 1

JournalClub
Editor’sNote:Theseshort,criticalreviewsofrecentpapersintheJournal,writtenexclusivelybygraduatestudentsorpostdoctoral
fellows,areintendedtosummarizetheimportantfindingsofthepaperandprovideadditionalinsightandcommentary.Formore
informationontheformatandpurposeoftheJournalClub,pleaseseehttp://www.jneurosci.org/misc/ifa_features.shtml.
HumanLearningImprovesMachineLearning:Neuraland
ComputationalMechanismsofPerceptualTraining
JoshuaCarpandJoonkooPark
DepartmentofPsychology,UniversityofMichigan,AnnArbor,Michigan48109
Review of Zhang et al.
Visual perception is highly plastic; per-
ceptual training dramatically increases
sensitivity to contrast, contour, and mo-
tion. However, many questions remain
about the neural mechanisms underlying
suchlearning-relatedchange.Previousre-
search shows that practice can increase
neuralselectivityfortrainedstimuli(Jiang
et al., 2007). Such selectivity might reflect
increased neural responses to preferred
stimuli, decreased responses to nonpre-
ferred stimuli, or a combination of the
two. These issues have both theoretical
and applied significance; improved un-
derstanding of the mechanisms support-
ing perceptual learning could lead to
important advances in the explanation
andtreatmentofvisualdisorders,includ-
ing amblyopia and age-related visual
impairment.
Zhang et al. (2010) used multivoxel
pattern analysis (MVPA) to explore the
neural mechanisms underlying practice-
related improvement in the perception of
visual form. MVPA uses fine-scale pat-
ternsofneuralactivation,measuredusing
functionalMRI,toclassifyobservers’per-
ceptual choices. This method has been
used to decode edge orientation and mo-
tion direction from brain responses to vi-
sual stimuli (Kamitani and Tong, 2005).
Zhang et al. (2010) measured neural acti-
vation in a form perception task before
and after psychophysical training in hu-
man subjects. Observers viewed patterns
that were parametrically varied along a
spectrum from radial to concentric forms
(Zhang et al., 2010, their Fig. 1A). The
authors then used MVPA to classify pat-
terns of neural activity according to the
parameters of the form stimulus that was
presented on each trial. To assess the ac-
curacy of these classifications, they calcu-
lated the likelihood of misclassification as
a function of the similarity between the
presented stimulus and the classifier’s
prediction.
Perceptual training improved classifi-
cation performance based on activity in
ventral and dorsal visual areas in two
ways.First,trainingincreasedthenumber
of stimuli classified correctly by MVPA
(i.e., the amplitude of the tuning func-
tion).Zhangandcolleagues(2010)attrib-
uted this result to enhanced blood
oxygenation level-dependent responses
across voxels that encode the preferred
stimulus category. Second, training de-
creased the number of grossly misclassi-
fied patterns (i.e., the SD of the tuning
function): when the classifier predicted
visual form incorrectly, the predicted
formtendedtobemoresimilartothetrue
form posttraining (Zhang et al., 2010,
their Fig. 2). The authors attributed this
result to decreased neural responses to
nonpreferred stimuli. The results also ex-
tend our knowledge about the role of the
dorsal visual stream in form perception.
Previous research has principally linked
form perception with the ventral stream,
or the “what” pathway; Zhang and col-
leagues (2010) show that the dorsal
“where” pathway contributes to this pro-
cess as well.
Zhang and colleagues (2010) showed
that training improves both behavioral
performance and neural selectivity in a
form perception task. These intriguing
results raise many questions about the
mechanismsunderlyingperceptuallearn-
ing. Successful task performance relies on
a series of computational processes: ob-
servers must detect and identify stimulus
input, apply decision rules, and maintain
attention to the task. Which of these pro-
cesses change during perceptual learning?
ThetrainingeffectsobservedbyZhang
et al. (2010) may reflect bottom-up
changes in the perception of visual form.
For example, perceptual learning might
improve observers’ ability to detect visual
stimuli, to integrate contrast information
acrosslocalorglobalscales,ortosegment
signal dots from noise. Alternatively,
learningmightaffecttop-downcontrolof
visual perception. Task performance re-
lies on decision processes that compare
theoutputsofcompetingpoolsofsensory
neurons (Heekeren et al., 2004). Observ-
ers must also maintain focus on the task;
performance declines during lapses of at-
tention (Christoff et al., 2009). Training
might improve these abilities as well. For
example, practice-related improvements
in form perception in the present study
might reflect gains in the ability to avoid
mind-wandering or lapses of attention.
ReceivedDec.20,2010;revisedJan.13,2011;acceptedJan.19,2011.
Correspondence should be addressed to Joshua Carp, Department
of Psychology, University of Michigan, 530 Church Street, Ann Arbor,
MI48109.E-mail:jmcarp@umich.edu.
DOI:10.1523/JNEUROSCI.6644-10.2011
Copyright©2011theauthors 0270-6474/11/313937-02$15.00/0
TheJournalofNeuroscience,March16,2011 • 31(11):3937–3938 • 3937

Page 2

Future experiments should parse the
contributions of bottom-up and top-
down processes to the training effects
documented by Zhang et al. (2010). Top-
down mechanisms related to perceptual
decision-making and attentional control
arethoughttorelyonprefrontalrepresen-
tations (Heekeren et al., 2004). If learning
improves top-down modulation of per-
ception, then training might alter neural
responses in prefrontal as well as visual
areas. In contrast, if learning sharpens the
tuning of neural-form representations via
bottom-up mechanisms, then learning-
related effects on neural tuning should
persist in the absence of top-down con-
trol—for example, when form stimuli are
presentedoutsidethefocusofattentionor
outsideconsciousawareness.Indeed,psy-
chophysical studies show that perceptual
learning can occur even for subconsciously
presented stimuli (Watanabe et al., 2001),
reflecting changes in bottom-up mecha-
nisms. However, if Zhang et al.’s (2010) re-
sultsprincipallyreflectchangesintop-down
processes, then the benefits of training
should only emerge when observers attend
toformstimuli.
Thepresentresultsalsoraisequestions
about the nature of the information ex-
tracted by MVPA: did the SVM classifier
usedinthepresentexperimentdetectrep-
resentations of objective information in
thevisualdisplayorsubjectiveinterpreta-
tions of that information? Serences and
Boynton (2007) showed that activation
patterns in both early and late visual areas
classified the true direction of visual mo-
tion stimuli. In contrast, only late visual
activation classified observers’ subjective
perception of motion independent of vi-
sual input. Future studies should extend
these results to the form perception task
used in Zhang et al. (2010), assessing the
contribution of each experience-sensitive
visual area to objective versus subjective
coding.
Zhang and colleagues (2010) showed
that perceptual training simultaneously
enhances neural responses to preferred
stimuli and suppresses responses to non-
preferred stimuli. These two observations
might reflect a single underlying phe-
nomenon. For example, learning might
directly increase neural responses to
preferred-form stimuli. This increased
activation might, in turn, decrease re-
sponses in neurons that represent non-
preferred stimuli via local inhibitory
circuits. Alternatively, enhancement and
suppression might stem from indepen-
dent mechanisms, with dissociable effects
onperceptionanddecisionmaking.Stud-
ies of the time course of perceptual train-
ing over multiple sessions may elucidate
the relationship between these phenom-
ena. Do enhancement and suppression
develop in parallel, or does one effect lag
theotherintime?Arethetwophenomena
correlated across trials, sessions, or sub-
jects? How does each effect relate to
learning-related improvements in be-
havioral performance?
Zhang et al.’s (2010) results also raise
interesting questions about the durability
and generality of training effects on be-
havioral performance and brain function.
Observers were scanned before and after
three psychophysical training sessions.
Learning-relatedimprovementinmotion
perception may last 10 weeks or longer
beyond the end of practice (Ball and
Sekuler, 1982). Neural and behavioral
correlates of form-perception training
may be similarly long-lived; alternatively,
such effects may decay quickly in the ab-
sence of continued practice. Future stud-
ies should also determine the extent to
which the effects of visual-form training
generalize to other tasks. For example, if
the learning-related effects observed by
Zhang and colleagues (2010) reflect gen-
eral improvements in form perception,
thistrainingprotocolshouldalsoimprove
the perception of contrast- and motion-
defined form, as well as the luminance-
defined forms used here. Training might
alsoaffecttherepresentationofnaturalis-
tic visual forms, including faces and
scenes. In contrast, if this protocol im-
proves form tuning by reducing mind-
wandering or increasing motivation to
attendtothetask,trainingmightimprove
performance across nearly any situation
that requires observers to sustain atten-
tion to a dull or repetitive task. Thus, fu-
ture studies should assess transfer effects
from the present training protocol to a
broad range of perceptual tasks.
Finally,theseresultsillustratethediffi-
culty of interpreting the sophisticated
techniques increasingly used in modern
neuroimaging. The present experiment
submittedtherawdatatoseveralstagesof
analysis. The researchers first submitted
activation estimates to 15 pairwise SVM
classifiers. Next, they aggregated these
two-way classifiers into a single six-way
classifier. Finally, they fit Gaussian tuning
curves to the output of the six-way classi-
fier. The complexity of this procedure
obfuscates the relationships between
successive stages of analysis: the output of
one stage may differ from that of the next
in unexpected ways. For example, al-
though the accuracy of the six-way classi-
fiercloselyparalleledtheamplitudeofthe
fitted tuning function in most regions,
these estimates yielded qualitatively dif-
ferent results in area V4v. Before fitting,
classification accuracy was higher for pre-
training than posttraining; after fitting,
thispatternwasreversed(compareZhang
et al., 2010, their Figs. 3A and S3). Simi-
larly,fittingreducedtheconfidenceinter-
vals of tuning amplitude by an order of
magnitude relative to classification accu-
racy. In summation, such multilayered
analytic techniques can substantially
changethemeaningofthedata.Research-
ersshouldapplyandinterpretthesemeth-
ods with caution.
References
BallK,SekulerR (1982) Aspecificandenduring
improvement in visual motion discrimina-
tion. Science 218:697–698.
Christoff K, Gordon AM, Smallwood J, Smith R,
Schooler JW (2009) Experience sampling
during fMRI reveals default network and ex-
ecutivesystemcontributionstomindwander-
ing.ProcNatlAcadSciU S A106:8719–8724.
Heekeren HR, Marrett S, Bandettini PA, Unger-
leider LG (2004) A general mechanism for
perceptual decision-making in the human
brain. Nature 431:859–862.
JiangX,BradleyE,RiniRA,ZeffiroT,VanmeterJ,
RiesenhuberM (2007) Categorizationtrain-
ing results in shape- and category-selective
human neural plasticity. Neuron 53:891–903.
Kamitani Y, Tong F (2005) Decoding the visual
and subjective contents of the human brain.
Nat Neurosci 8:679–685.
Serences JT, Boynton GM (2007) The representa-
tion of behavioral choice for motion in human
visualcortex.JNeurosci27:12893–12899.
Watanabe T, Na ´n ˜ez JE, Sasaki Y (2001) Percep-
tual learning without perception. Nature
413:844–848.
Zhang J, Meeson A, Welchman AE, Kourtzi Z
(2010) Learning alters the tuning of func-
tional magnetic resonance imaging patterns
for visual forms. J Neurosci 30:14127–14133.
3938 • J.Neurosci.,March16,2011 • 31(11):3937–3938 CarpandPark•JournalClub