Content uploaded by John Michael Pearson
Author content
All content in this area was uploaded by John Michael Pearson on Oct 27, 2014
Content may be subject to copyright.
NATURE NEUROSCIENCE volume 15 | number 3 | mArCH 2012 341
NEWS AND VIEWS
inferences about the underlying decision algo-
rithm. Typical studies only detect correlates
of specific model-inspired variables, but tem-
poral patterns of activation are, in principle,
much more distinct, implying greater power
to compare models. In fact, the authors were
able to compare several alternative decision
models and found only one among their set
that adequately reproduced task-dependent
activity in the brain.
In the authors’ experiment, participants
chose between pairs of options (rewarded
with points that were eventually exchanged
for money). For each option, participants
were shown both the number of points they
might get and the probability with which they
would get them. Thus, subjects had to con-
sider both pieces of information before mak-
ing a decision: a higher point-value option
with low probability may well be the worse
of the two options. Hunt et al.1 meas ured the
time it took participants’ to make their deci-
sions, which depended on both the sum and
difference in the values of the two options.
To model the network dynamics underlying
the decision, Hunt et al.1 employed a simpli-
fied version of previously described reverber-
ating networks2–5. In these models, mutually
inhibitory pools of neurons represent each
option and implement a winner-take-all
competition that represents the outcome
of the decision process. As a result, higher
overall values across all options lead to higher
neuronal activation, and thus to more total
inhibition, reducing decision times, whereas
small differences in option values lead to
more difficult decisions, less inhibition and
longer reaction times6,7. Most model param-
eters are constrained biophysically, leaving a
few parameters, such as the speed-accuracy
tradeoff, to be fit to individual participant’s
choices. By identifying total synaptic activity
in the model with the measured MEG signal,
NEWS AND VIEWS
The authors are in the Department of Neurobiology,
Duke University Medical Center, Duke Institute for
Brain Sciences, Durham, North Carolina, USA.
e-mail: platt@neuro.duke.edu
Dynamic decision making in the brain
John Pearson & Michael L Platt
How do we make decisions? A study uses MEG to provide the spatial as well as the temporal resolution needed to
answer this question, together with computational modeling, which allows for complex non-linear decision models.
This work helps resolve some of the seemingly contradictory results from previous work.
What happens when we decide between two
different routes on our morning commute
or two different meals at lunch? How do we
weigh costs and benefits against one another
and do so in a way that minimizes mental
effort and regret? For economists, psycholo-
gists and others who study decision making,
these questions form the heart of the science
of choice. For neuroscientists, such concerns
inform the study of neural circuit dynam-
ics that translate sensory inputs into behav-
ior. What we refer to as the decision occurs
between these inputs and behavioral outputs,
and any neurobiologically complete account
of this process must explain the flow of infor-
mation from one set of neurons to the next,
from retina to eye movement or cochlear hair
cell to button press.
Intuitively, the richness and flexibility of
such transformations must depend on the
multiplexed connections and feedback among
neurons in the brain, which result in a highly
interactive nonlinear system. In theory, differ-
ent types of decision algorithms should give
rise to distinct patterns of network activity,
thus yielding clues to the underlying neural
computations. In practice, most current
experimental techniques are severely limited
in their ability to examine network dynamics,
sacrificing whole-brain information for direct
access (as in individual neuron recordings) or
focusing on either temporal (electroencepha-
lograms) or spatial information (functional
magnetic resonance imaging, fMRI). In this
issue of Nature Neuroscience, Hunt et al.1 set
out to address this gap via a new, model-based
approach to studying whole-brain network
dynamics using magneto-encephalography
(MEG), which measures magnetic fields gener-
ated by electrical currents in the brain. Although
typical experiments use theoretical models to
supply putative correlates of hemodynamic
response or neuronal firing, Hunt et al.1 simu-
late a well-known decision algorithm to pre-
dict the measured neural signal itself (Fig. 1).
That is, by identifying correspondences
between their model’s total synaptic input and
the physiological MEG signal, Hunt et al.1 we re
able to look for correspondences between mea-
sured data and their simulated brain.
This study makes several key advances
over previous studies. First, the high tempo-
ral resolution of MEG, in combination with
source reconstruction, provides a whole-
brain picture of the unfolding decision pro-
cess with millisecond precision. Task-related
activations begin in visual cortex, spread to
fronto-polar and ventromedial prefrontal
regions, pass to medial and lateral parietal
cortex, and conclude in motor areas at the
time participants press the response button,
all in a matter of a few seconds.
Second, by modeling measured neural
responses directly, Hunt et al.1 were able to use
standard linear methods to study highly non-
linear decision models. That is, rather than
attempt to correlate model-based variables
with MEG responses, the authors produced
two parallel datasets—real and simulated
MEG, broken into distinct frequency bands—
and subjected both to the same (linear)
correlation analyses (Fig. 1). By looking for
correspondences between the two sets of
results, the authors were able to search for
telltale signs of the decision algorithm with-
out needing to invert the complex relation-
ship between model inputs and MEG.
Finally, by using a model-based pre-
diction for the time course of MEG data,
Hunt et al.1 were able to make compelling
npg © 2012 Nature America, Inc. All rights reserved.
342 volume 15 | number 3 | mArCH 2012 NATURE NEUROSCIENCE
NEWS AND VIEWSNEWS AND VIEWSNEWS AND VIEWS
behavior as the output of the brain’s highly
dynamic networks. By allowing models to
motivate not only the analysis, but also the
design, of experiments, it also generates new
opportunities for theoretical neuroscience.
As such, it stands not only to enhance the
quality of our collected data, but the power
and realism of our models.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
1. Hunt, L. et al. Nat. Neurosci. 15, 470–476 (2012).
2. Soltani, A. & Wang, X.J. J. Neurosci. 26, 3731–3744
(2006).
3. Wang, X.J. Neuron 36, 955–968 (2002).
4. Wong, K.F., Huk, A.C., Shadlen, M.N. & Wang, X.J.
Front. Comput. Neurosci. 1, 6 (2007).
5. Wong, K.F. & Wang, X.J. J. Neurosci. 26, 1314–1328
(2006).
6. Shadlen, M.N. & Newsome, W.T. J. Neurophysiol. 86,
1916–1936 (2001).
7. Usher, M. & McClelland, J.L. Psychol. Rev. 108,
550–592 (2001).
8. Kable, J.W. & Glimcher, P.W. Nat. Neurosci. 10,
1625–1633 (2007).
9. Plassmann, H., O’Doherty, J. & Rangel, A. J. Neurosci. 27,
9984–9988 (2007).
10. Blair, K. et al. J. Neurosci. 26, 11379–11386
(2006).
11. Platt, M.L. & Glimcher, P.W. Nature 400, 233–238
(1999).
12. Sugrue, L.P., Corrado, G.S. & Newsome, W.T. Science 304,
1782–1787 (2004).
13. Kable, J.W. & Glimcher, P.W. Neuron 63, 733–745
(2009).
14. Noonan, M.P. et al. Proc. Natl. Acad. Sci. USA 107,
20547 (2010).
15. Rorie, A.E., Gao, J., McClelland, J.L. & Newsome, W.T.
PLoS ONE 5, e9308 (2010).
Hunt et al.1 were then able to make telltale
predictions for the neural signature of the
winner-take-all decision.
The authors found that activity in the ven-
tromedial prefrontal cortex (vmPFC) and
the posterior superior parietal lobule (pSPL)
strongly followed the predictions of the
model. In both areas, overall value exerted a
strong effect in the 2–10-Hz frequency range
early in the trial, whereas value difference
dominated later on, with a response confined
to lower frequencies (2–4 Hz). Although the
particular locations of these effects might
have been anticipated on the basis of previ-
ous studies, in this case, the observed pattern
provides evidence for a particular computa-
tion underlying the decision process8–12.
Moreover, these results may help to resolve a
dispute over whether vmPFC encodes over-
all value or value difference13,14. Hunt et al.1
found representations of both quantities,
each with a distinct time course, in a duration
that was too short to be distinguishable by
fMRI. Moreover, these patterns are consistent
with previous recordings from single neurons
in monkeys15.
In addition, activity in these areas bore out
two key predictions of the decision model.
First, although correct trials showed an effect
of overall value early in the trial and of value
difference later in the trial, this value differ-
ence signal was blunted on error trials, as
would be expected if the information were
poorly encoded or weakly transmitted to these
areas. Second, participants’ individual speed-
accuracy tradeoffs, as reflected in median
reaction time, predicted variability in the cou-
pling of their MEG signals to the overall value
of both options. Together, these results lend
additional weight to the finding that vmPFC
and pSPL contribute to the implementation
of a winner-take-all decision.
Despite these findings, several questions
remain to be answered. Although Hunt et al.1
were able to show that their results both were
robust to parameter changes in the biophysi-
cal model and provided a better account of
the observed MEG data than other commonly
used models, they were not able to rule out
alternative networks that might also receive
value-related information and interact with
the vmPFC and pSPL. That is, other brain
areas may be involved in the decision pro-
cess in ways that are difficult to describe with
simple models. Moreover, it may be possible
that, for other tasks, very different single-unit
computations give rise to similar large-scale
signals, rendering it difficult to distinguish
between some classes of models. Finally, MEG
has limited utility in measuring activation in
subcortical brain areas that are thought to
substantially contribute to decision making3.
The model-based approach of Hunt et al.1
opens new possibilities for investigating
Time
Synaptic activity
Neural network Predicted signal
MEG signal
Source
reconstruction
Localized brain activity
Simulation
Band-pass
filter Wavelet
transform
Correlate
with task
variables
10
Time-frequency
significance map
Time-frequency
power spectrum
6
20 400 800 1,200
Time (ms)
Frequency (Hz)
10
6
20 500 1,000
Time (ms)
Frequency (Hz)
Figure 1 Schematic for analysis of decision-making dynamics in the brain. Simulated data from a winner-take-all neural network model are used to produce
predictions of synaptic activity to be compared to responses of individual brain regions measured with source-reconstructed MEG. Both real and
predicted signals are then filtered and wavelet transformed to plot brain activity across time and frequency. Finally, activity in regions of interest is
correlated with task-specific variables to produce maps of statistical significance. Warmer colors indicate a more significant effect of task variables at a
given time in a given frequency band.
npg © 2012 Nature America, Inc. All rights reserved.
