Available online at www.sciencedirect.
com
Computational models of decision making: integration, stability,
and noise
Nicholas Cain and Eric Shea-Brown
Decision making demands the accumulation of sensory
evidence over time. Questions remain about how this occurs,
but recent years have seen progress on several fronts. The first
concerns when optimal accumulation of evidence coincides
with the simplest method of accumulating neural activity:
summation over time. The second involves what computations
the brain might perform when summation is difficult due to
imprecision in neural circuits or is suboptimal due to uncertainty
or variability in how evidence arrives. Finally, the third concerns
sources of noise in decision circuits. Empirical studies have
better constrained the extent of this noise, and modeling work
is helping to clarify its possible origins.
Address
Department of Applied Mathematics, Program in Neurobiology and
Behavior, University of Washington, Seattle, WA, USA
Corresponding author: Shea-Brown, Eric (etsb@amath.washington.edu)
Current Opinion in Neurobiology 2012, 22:1047–1053
This review comes from a themed issue on Decision making
Edited by Kenji Doya and Michael N Shadlen
For a complete overview see the Issue and the Editorial
Available online 15th May 2012
0959-4388/$ – see front matter, # 2012 Elsevier Ltd. All rights
reserved.
http://dx.doi.org/10.1016/j.conb.2012.04.013
Introduction
For many sensory discrimination tasks, instantaneous
snapshots of sensory input represent single, noisy samples
that alone carry limited evidence. A prototypical example
is the random dots task, in which the subject must deter-
mine the average direction of motion in a field of ran-
domly moving dots (Figure 1). Here, as in many more
complex decisions, a sound judgment requires accumu-
lating this sensory evidence over time [1–7].
We review recent progress on how this type of accumu-
lation could occur in the brain. We begin with how the
neural encoding of this evidence determines optimal
mechanisms for integrating it over time. Next, we discuss
the stability and variability of neural circuits that may
perform this computation, the function and adaptability
of these circuits in uncertain decision making environ-
ments, and how they might operate with irregular and (at
least apparently) noisy components.
www.sciencedirect.com
Integration, optimality, and functionality
What is the optimal way to accumulate evidence over
time? Many different definitions of ‘optimal’ — from
maximizing the rate at which correct responses are made
[8,1] to maximizing guaranteed payoff [9] to lossless
Bayesian decoding [10] — depend on the same key
computation. This is to update the statistical likelihood
of each decision alternative being correct, relative to the
other alternatives, when each increment of evidence
arrives [8,11,12,13]. In general this is a tall order, requir-
ing computation of an arbitrary function of the momen-
tary neural activity that encodes incoming evidence, and
then combining this with the evidence already accumu-
lated. Making matters worse, even knowing what func-
tion to compute depends on full knowledge of the
statistics of that activity [11].
Fortunately, in a number of circumstances, the likelihood
calculation is equivalent to directly summing, or integrat-
ing, the evidence stream over time (see Figure 1). In the
simplest case there are two task alternatives, and two
independent streams of incoming evidence I1(t) and
I2(t) — for example, representing the output of neurons
selective for upward vs. downward motion. Perhaps the
best known result says that, if I1(t) and I2(t) are uncorre-
lated Gaussian signals with identical variance, then the
relative likelihood of each alternative can be computed by
integrating the signals over time — [8,11,12,13], via the
accumulated inputs
Z t
A1 ðtÞ ¼ I 1 ðt 0 Þdt 0 (1)
0
Z t
A2 ðtÞ ¼ I 2 ðt 0 Þdt 0 (2)
0
If I1(t) and I2(t) are Poisson processes — that is, spike
trains [10,14,15,16] — the optimal computation is
again integration over time. In fact, this setting is much
more flexible, as there is no longer a requirement that all
inputs have the same variance ([15,16] identify limits in
which the Poisson and Gaussian cases agree). Moreover,
for decisions involving multiple alternatives, where
motion could follow any of the directions of the compass
rose, the same result extends. In this case, there will be N
different input streams Ii(t), whose integrals produce N
accumulated variables Ai(t). Additionally, for the two
alternative cases, [17] extends these results to compute
estimates of the quality of evidence streams for each task
alternative.
Current Opinion in Neurobiology 2012, 22:1047–1053
1048 Decision making
Figure 1
time
tim
e
time
Schematic of sensory evidence accumulation over time. In the random dots task, a field of irregularly moving dots is presented to the subject, with
motion biased toward one favored direction. While some neurons represent ‘instantaneous’ values of this motion evidence (top rastergram), other
neurons appear to represent its accumulated value over time (bottom rastergram). According to the equation shown, the relationship between the
average activity of these neural pools approximates temporal integration.
When the statistics of incoming evidence vary over the
course of a trial, in some cases the relative likelihood of
task alternatives can still be computed based on the
integral of the inputs. For two Gaussian input streams,
this holds when the ratio of the ‘signal’ hI1(t)  I2(t)i in
the inputs to the variance of these inputs is constant over
time [18]. For N Poisson input streams, there are stronger
results: if the time dependence is via a gain term that
uniformly scales up or scales down the firing rate of each
input stream, then integration once again enables a direct
computation of the relative likelihoods; similar consider-
ations hold for variation across trials [10].
Can we measure such integration occurring in neural
circuits? Compelling evidence is presented in [19],
where the authors devise a novel method to measure
the time-evolution of the variance of spike rates believed
to represent the integrals Aj(t). The key finding is that this
variance grows linearly in time, consistent with what is
expected from an integrating process.
Nevertheless, computing an integral is not the end of the
story: integrated evidence must trigger an explicit
decision. For the two alternative case, the highest
accuracy at any given speed is achieved by making a
decision when the difference A1(t)  A2(t) crosses a preset
bound [13,8,11,20]. (We note that this still leaves open
precisely what tradeoff between speed and accuracy the
decision maker will select, which is a distinct and inter-
esting question [21–23,9,1,8].) For multiple alternatives,
Ai(t) can be used to implement a ‘MSPRT’ rule which can
also optimize speed and accuracy [24,16], or to find
the maximum likelihood alternatives at any given time
[10]. While in principle these decision rules could be
Current Opinion in Neurobiology 2012, 22:1047–1053
I1(t)
time
A1(t) ≈ I1(t) dt
A1(t)
N
time
Current Opinion in Neurobiology
computed in ‘downstream’ areas, recurrent inhibition
within the same networks that integrate the evidence
streams Ij(t) can often implement or closely approximate
the required differencing operations [7,18,25,26,1].
Beyond the integral sign: precision,
thresholding, and the neural ratchet
While ‘pure’ integration of incoming evidence is optimal
in many settings, it poses a challenge for neural circuitry.
Evidence is integrated over hundreds of milliseconds or
longer as decisions develop, but the activity of individual
neurons and synapses tend to decay with timescales that
are several orders of magnitude faster. How can this rapid
decay be countered? A classical solution is via feedback
connections tuned to balance — and hence cancel —
passive voltage leak and synaptic decay [27,30]. This
process is illustrated in Figure 2a via motion of a ball on an
energy surface. The position A(t) represents the total
activity of a circuit (say, average firing rate relative to a
baseline marked 0). If decay dominates (upper-right),
then A(t) always has a tendency to ‘roll back’ to baseline
values; we say the integrator circuit is stable. Conversely, if
feedback connections are in excess, then activity will
grow away from the baseline value and thus the circuit
is unstable. Pure integration is achieved (upper-left) with
feedback that is precisely tuned to match decay, so that
changes in A(t) represent integrated evidence alone
[27,31].
Biologically, this feedback could occur in many forms.
Beyond excitatory connections, positive feedback can also
occur through opponent inhibition [32,1,33,27,28,7,34];
recently, a generically valid ‘normal form’ equation that
describes the resulting dynamics was presented and its
www.sciencedirect.com
 Computational models of decision making Cain and Shea-Brown 1049

Figure 2

(a) Integration Instability Stability (b)

I(t) ?
Photon absorption: accumulate over space

A(t) A(t) A(t)

0 0 0

A(t) A(t) A(t) Motion direction: accumulate over time

0 0 0
Current Opinion in Neurobiology

(a) Visualizing integration via an energy surface [27,28,29], where the location of the ball indicates integrated evidence A(t). A robust integrator can
‘fixate’ at a range of discrete values, indicated by a sequence of potential wells, despite imprecise tuning of circuit feedback. Without these wells (the
non-robust case), activity in a mistuned integrator would either exponentially grow or decay, as in the top panels. Perturbing the robust integrator from
one well to the next, however, requires sufficiently strong momentary input I(t). (b) Relationship of robust integration to the detection of sparse visual
signals. Top: activity of five photoreceptors is shown as a dot, and the dotted line represents photon absorption. Light that arrives at only a tiny fraction
of photoreceptors (here receptor # 3) must be detected. Bottom: motion evidence fluctuates over time, and the direction of the time-average must be
discriminated among two possibilities — that is, up (dotted line) vs. down (solid line).

predictions favorably compared with behavioral data [35].
Interestingly, unstructured networks also appear able to
accumulate inputs over time: a single ‘integration mode’
may emerge naturally as the connection statistics are varied
in large, randomly coupled networks [36]. Recent studies
have also raised the possibility of achieving integration
without depending on feedback per se, rather by passing
inputs through extended sequences of ‘functionally feed-
forward’ states within a large network [37–39]. This mech-
anism is one of several that produce high-dimensional
dynamics — that is, with time constants that differ from
cell to cell — by contrast to classical line attractor models.
Intriguingly, such cell-to-cell heterogeneity has recently
been found in the zebrafish oculomotor integrator circuit
[40]. However they are implemented, must feedback
effects always be precisely tuned to produce integration
over time? The robust integrator mechanism of [41] presents
an alternative. Here, the circuit as a whole ‘ratchets’ among
many stable states, as illustrated in the bottom row of
Figure 2a by introducing a series of wells into the energy
surface [27,28,29]. Importantly, even with imprecisely
tuned feedback, the network now avoids the tendency
to intrinsically decay or grow without incoming evidence.
Thus, the activity A(t) approximates the integral of I(t),
without the effects of decay or runaway excitation.
Nevertheless, such robustness impacts how the evidence
streams Ij(t) are processed. As the energy wells illustrate,
instantaneous inputs Ij(t) below a threshold value will fail
to perturb the state from one well to another [41,42], and
are therefore ignored by the integrator. The idea of
thresholding inputs has a long history [43,44], and several
studies explore its implications for decision making over
time. In [45], the authors show that a closely related
‘interval of uncertainty’ model can match distributions
of reaction times and correct vs. incorrect responses better
than an allied model without a threshold. Moreover, [46]
develops a model with a threshold-like ‘gating’ effect,
finding improved fits to single unit recordings while
simultaneously matching behavioral data (see also
[47]). Several key statistical and dynamical properties
of robust integrators that contribute to the quality of this
match are studied in [48]. Ongoing work asks how robust
integration impacts the optimality of decision perform-
ance. For example, consider the case introduced above
where instantaneous evidence is carried by Gaussian
signals I1(t) and I2(t). Rather than accumulate the entire
signal, a robust integrator would ignore all samples below
a preset threshold. Intuitively, one might expect this to
diminish the speed and accuracy with which decisions can
be made. However, results to date [49] suggest that the
loss is minimal, even when more than half of the signal is
thresholded away. This suggests that most of the evi-
dence can be gleaned from the ‘tails’ of the signals.
Moving beyond the case of Gaussian inputs, analogy with
other fields suggests some settings in which thresholding
actually improves decision performance. For example,
consider the case of detecting photons that arrive rarely
at noisy photoreceptors (see Figure 2b). If the output of
all photoreceptors is simply summed — that is, inte-
grated — this sparse signal could easily be lost in noise.
The solution demands summing receptor responses only
after passing through a thresholding function [43], analo-
gous to integrating only those inputs strong enough to
move the robust integrator state from well to well. For
decisions based on accumulating temporally sparse evi-
dence over a noisy background — such as the auditory
and visual impulses recently studied in rats and humans
[50] — thresholding integrators would then be expected

www.sciencedirect.com Current Opinion in Neurobiology 2012, 22:1047–1053

1050 Decision making
to improve decision performance. Taken together, these
observations point to a potentially favorable role for
robust integrators in decision making. Ignoring the
weaker parts of the input signal is sometimes the best
strategy, and when it is not, the cost of doing so can be
surprisingly low.
Richer computations over time
Beyond sparseness, there are other conditions that also
demand circuit dynamics departing from pure integration.
For example, consider a version of the moving dots task in
which the stimulus (i.e. mostly upward or downward
moving dots) is prefaced by a period of purely random
motion. Then, a simple integral of the inputs would
produce a mix of both noise and signal-carrying inputs.
A better strategy is to put more emphasis on stimulus
periods that contain the signal.
If the stimulus onset is explicitly cued, or accompanied by
a large sensory change, then there are several straightfor-
ward ways to accomplish this. For example, the integrator
circuit could be ‘reset’ when the stimulus occurs [51,52],
erasing integrated noise samples. A recent physiology
study provides support for this notion [19]: spike rate
variance in an area believed to represent integrated
evidence drops strongly after stimulus onset. Beyond
the reset, other studies have asked what types of stable
and unstable circuit dynamics could optimally process
such time-dependent stimuli (see Figure 2a). Not surpris-
ingly, optimal solutions begin with stability in periods
before the stimulus arrives, so that the effect of noisy
inputs remains contained. Once the stimulus arrives, the
best strategy is to transition to either pure integration or
instability, thus letting evidence begin to build up over
time [7,18,53,39].
What mechanisms could drive changes in integrator
stability over time? One possibility is transient release
of a neuromodulator such as norepinephrine (NE). One
model of the impact of NE is that it controls the strength,
or gain, of connections within an integrator circuit [54]. If
NE were to transiently increase the gain of feedback, for
example, it could switch a circuit from being stable to
unstable. Network models have shown how this could
lead to improved decision performance [18,55,53,56].
Determining the timing of NE release, however,
represents a decision in itself — and this ‘first’ decision
about the presence of a stimulus can only be made in time
to improve the ‘main’ decision about the stimulus content
under specific assumptions about the decision circuit [53].
More recently, NE transients have been analyzed within
biologically detailed models of decision making [55].
The authors show how NE can steer a network between
stable, unstable, and integrating regimes, and develop
predictions for the density and location of NE receptors
that optimize decision performance. Moreover, they show
Current Opinion in Neurobiology 2012, 22:1047–1053
how the local composition of NE receptors determines
which regime occurs in a given circuit. This opens the
door to circuit-specific effects of diffuse NE modulation
[55] — a specificity that might be of use when distinct
evidence streams must be simultaneously integrated over
time, as has been recently demonstrated to occur [50].
Related studies have shown how circuit dynamics should
differ from pure integration when other aspects of the
decision making environment are variable and uncertain.
If the difficulty of a task varies from trial to trial — as in
many experimental settings [12,52] — then an ‘urgency
signal’ may be added to integrated evidence [51]. These
authors note that such a signal is equivalent to a gradual
collapse of the threshold for making a decision. This
results in truncation of excessively long decisions,
thereby concentrating net effort toward easier trials—a
strategy which can provide optimal decision performance
[57–60]. Models with time-dependent gain mechanisms
[61], or with instabilities in the dynamics that integrate
evidence [62], instantiate a similar effect. These results
may find synergy with the concept that instability and
gain are modulated to adjust to a related type of uncer-
tainty, in which the relationship among stimuli, decisions,
and rewards is not fully known. Here, different levels of
stability — possibly mediated by NE — could differently
emphasize conservative, gradual accumulation of inputs
vs. rapid, runaway commitment to ‘exploratory’ decisions
based on chance fluctuations in evidence streams
[54,55,56], cf. [63].
Integration, background, and circuit-driven
noise
Above, we have treated the transformation of the incom-
ing evidence I(t) into its integral A(t) as a deterministic
mathematical operation. However, A(t) is presumably
computed by irregularly spiking neurons (Figure 1, bot-
tom raster), perhaps distributed throughout the brain —
and possibly incorporating ‘background’ neurons that are
untuned or poorly tuned to a specific decision. Do inte-
grator circuits with these noisy components contribute
additional variability to decisions?
Empirical work suggests two constraints. First, well-
established patterns of accuracy and reaction time distri-
butions [12,52,11] must be compatible with whatever
type of circuit-based noise arises. Second, for the moving
dots task, a number of studies have precisely measured
the signal and noise properties of motion evidence at a
stage (area MT) that appears to precede its accumulation
over time [64,65,66,67]. Recently, building on prior
analyses, [66,68,69] showed that a noiseless accumu-
lation, or sum, of MT spikes over an empirically esti-
mated distribution of reaction times predicts levels of
decision performance that closely match experiment.
This only leaves room for a relatively small amount of
www.sciencedirect.com
 Computational models of decision making Cain and Shea-Brown 1051
additional variability to be contributed by the integrator
circuit itself.
Spiking neuron models of integrator circuits have met
some of these constraints. A number of studies show that
such circuits can reproduce behavioral statistics as well as
key features of neural activity [10,70,68,71,34]. In
addition, [10] shows that additional properties of noise-
less integration — including how confidence in task altern-
atives grows over time — are preserved in a simple
integrator circuit with Poisson-like firing (see also
[68,72]). For biologically detailed models — especially
those in which the long time constants of integration arise
through recurrent network interactions [34,73,70,74] —
future work will be needed to understand how much
variability arises from the fluctuating inputs I(t), from
background neurons, and from the irregular dynamics of
the circuits themselves.
Three recent advances in theoretical neuroscience lay
important groundwork. The first is the model of [71], in
which past evidence is represented by spikes that recur at
randomly spaced times, as in a Poisson process — allow-
ing them to continue to contribute variability to the
decision process. While this is an abstract model, it does
separate the variability driven by the evidence stream and
that generated by mechanisms within the integrator cir-
cuit itself, a perspective that will remain important in
future studies.
The second, related development concerns how integra-
tor circuits process both noise in incoming signals and
noise local to each ‘cell’ in the circuit [39,37,75]. In
particular, functionally feedforward architectures amplify
incoming signals faster than internally occurring noise.
This could clearly improve decision making performance,
although constraints such as saturation of firing rates may
limit when this strategy can be applied [39,37]. Finally,
[76] shows how lossless statistical inference over time can
be performed by the (integrate-and-fire) dynamics of
spiking neural networks. Importantly, this study shows
how irregular, variable spiking and optimal neural integ-
ration can coexist, and suggests that — at least in prin-
ciple — all of the truly detrimental noise in integrator
circuits could be sourced to variability upstream.
Conclusion
Models of decision making have an enduring theoretical
foundation in the accumulation of evidence streams over
time. In more and more settings, we have learned that
this computation can be implemented by simply inte-
grating neural activity over time. We have also learned
about when and why we need to reach beyond this
concept. For example, exact integration over long time-
scales may be difficult to achieve in realistic circuits.
However, this apparent ‘bug’ — and the richly varied
circuit dynamics that could result — may become a
www.sciencedirect.com
‘feature’ in unpredictable decision making environ-
ments. Adding to the possibilities, different neural inte-
grator circuits vary not only in how they process
fluctuating evidence, but also in whether and how they
may contribute additional noise to the decision compu-
tation. The future is bright for careful combinations of
experiment, modeling, and theory that will link increas-
ingly realistic circuit models with optimal algorithms for
decision making in uncertain environments.
Acknowledgements
We thank Rafal Bogacz, Anne Churchland, Philip Holmes, and the editors
for their valuable comments on the manuscript, and these and many other
colleagues for their contributions to our understanding of this material over
the years. We gratefully acknowledge the support of a Career Award at the
Scientific Interface from the Burroughs Welcome Fund (E. S. B.).
References
1. Bogacz R, Brown E, Holmes P, Cohen JD: The physics of optimal
decision making: a formal analysis of models of performance
in two-alternative forced-choice tasks. Psychol Rev 2006,
113:700-765.
2. Laming DRJ: Information Theory of Choice-Reaction Times. New
York: Academic Press; 1968.
3. Ratcliff R: A theory of memory retrieval. Psych Rev 1978,
85:59-108.
4. Ratcliff R, Van Zandt T, McKoon G: Connectionist and diffusion
models of reaction time. Psych Rev 1999, 106:261-300.
5. Shadlen MN, Newsome WT: Motion perception: seeing and
deciding. PNAS 1996, 93:628-633.
6. Stone M: Models for choice-reaction time. Psychometrika 1960,
25:251-260.
7. Usher M, McClelland JL: The time course of perceptual choice:
the leaky, competing accumulator model. Psychol Rev 2001,
108:550-592.
8. Gold J, Shadlen M: Banburismus and the brain: decoding the
relationship between sensory stimuli, decisions, and reward.
Neuron 2002, 36:299-308.
9. Zacksenhouse M, Bogacz R, Holmes P: Robust versus optimal
strategies for two-alternative forced choice tasks. J Math
Psychol 2010, 54:230-246.
10. Beck JM, Ma WJ, Kiani R, Hanks T, Churchland AK, Roitman J,
 Shadlen MN, Latham PE, Pouget A: Probabilistic population
codes for Bayesian decision making. Neuron 2008,
60:1142-1152.
The authors show how Bayesian inference — that is, computing the
relative likelihood that alternative decisions will be correct — can be
performed by integrating spikes over time. In particular they show that
this holds for broad classes of firing statistics within the ‘exponential
family,’ which includes Poisson statistics, and for many assumptions on
how the stimulus is encoded within a neural population
11. Smith PL, Ratcliff R: Psychology and neurobiology of simple
decisions. Trends Neurosci 2004, 27:161-168.
12. Gold JI, Shadlen MN: The neural basis of decision making. Annu
 Rev Neurosci 2007, 30:535-574.
This review covers important results that relate classical sequential
analysis theory to both behavioral electrophysiological observations
13. Wald A, Wolfowitz J: Optimum character of the sequential
probability ratio test. Annals Math Statistics 1948, 19:326-339.
14. Link SW: The Wave Theory of Difference and Similarity. Lawrence
Erlbaum; 1992.
15. Smith PL: From Poisson shot noise to the integrated Ornstein-
Uhlenbeck process: neurally principled models of information
Current Opinion in Neurobiology 2012, 22:1047–1053
1052 Decision making
accumulation in decision-making and response time. J Math
Psychol 2010, 54:266-283.
16. Zhang J, Bogacz R: Optimal decision making on the basis of
 evidence represented in spike trains. Neural Comput 2010,
22:1113-1148.
This paper applies classical sequential sampling procedures to the
setting of neural population coding. This yields concrete and useful
results on when optimal algorithms for evidence accumulation can be
accomplished by integrating inputs over time — and when this will fail.
Results are given for decisions among two and multiple alternatives
17. Moreno-Bote R: Decision confidence and uncertainty in
diffusion models with partially correlated neuronal
integrators. Neural Comput 2010, 22:1786-1811.
18. Brown E, Gao J, Holmes P, Bogacz R, Gilzenrat M, Cohen J:
Simple networks that optimize decisions. Int J Bifurcation
Chaos 2005, 15:803-826.
19. Churchland AK, Kiani R, Chaudhuri R, Wang X-J, Pouget A,
 Shadlen MN: Variance as a signature of neural computations
during decision making. Neuron 2011, 69:818-831.
The authors develop a method to assign noise in single-cell responses to
one of two sources: (1) variability in an underlying stochastic process that
could represent the time integral of decision evidence, and (2) additional
‘point-process’ noise, such as that emitted from a neuron firing as a
Poisson process. This shows a ‘reset’ of the underlying variability to lower
values when task stimuli arrive, followed by a gradual increase of this
variability while these stimuli are integrated over time. Both of these
properties are consistent with optimal accumulation of evidence over
time
20. Kiani R, Hanks TD, Shadlen MN: Bounded integration in parietal
cortex underlies decisions even when viewing duration is
dictated by the environment. J Neurosci 2008, 28:3017-3029.
21. Bogacz R, Hu P, Holmes P, Cohen JD: Do humans produce the
speed-accuracy tradeoff that maximizes reward rate? Quart J
Exp Psych 2010, 63:863-891.
22. Edwards W: Optimal strategies for seeking information:
models for statistics, choice reaction times, and human
information processing. J Math Psychol 1965, 2:312-329.
23. Maddox WT, Bohil CJ: Base-rate and payoff effects in
multidimensional perceptual categorization. J Exp Psychol
1998, 24:1459-1482.
24. Bogacz R, Gurney K: The basal ganglia and cortex implement
optimal decision making between alternative actions. Neural
Comput 2007, 19:442-477.
25. McMillen T, Holmes P: The dynamics of choice among multiple
alternatives. J Math Psychol 2006, 50:30-57.
26. Wong KF, Wang XJ: A recurrent network mechanism of
time integration in perceptual decisions. J Neurosci 2006,
26:1314-1328.
27. Goldman MS, Compte A, Wang X-J: Neural integrator models.
 In: Squire LR, editor. Encyclopedia of Neuroscience, 6:165–178,
2009.
An excellent review of neural mechanisms for signal integration that
bridges fields of neuroscience
28. Machens CK, Romo R, Brody CD: Flexible dynamics of mutual
inhibition: a neural model of two-interval discrimination.
Science 2005, 307:1121-1124.
29. Pouget A, Latham P: Digitized neural networks: long-term
stability from forgetful neurons. Nat Neurosci 2002, 5:709-710.
30. Robinson DA: Integrating with neurons. Ann Rev Neurosci 1989,
12:33-45.
31. Seung HS: How the brain keeps the eyes still. Proc Natl Acad Sci
1996, 93:13339-13344.
32. Bogacz R: Optimal decision-making theories: linking
neurobiology with behavior. Trends Cogn Sci 2007,
11:118-125.
33. Brown E, Holmes P: Modeling a simple choice task: stochastic
dynamics of mutually inhibitory neural groups. Stochastics
Dynam 2001, 1:159-191.
Current Opinion in Neurobiology 2012, 22:1047–1053
34. Wang X-J: Probabilistic decision making by slow reverberation
in cortical circuits. Neuron 2002, 36:955-968.
35. Roxin A, Ledberg A: Neurobiological models of two-choice
decision making can be reduced to a one-dimensional nonlinear
diffusion equation. PLoS Comput Biol 2008, 4:e1000046.
36. Ganguli S, Bisley J, Roitman J, Shadlen M, Goldberg M, Miller K:
One-dimensional dynamics of attention and decision making
in lip. Neuron 2008, 58:15-28.
37. Ganguli S, Huh D, Sompolinsky H: Memory traces in dynamical
systems. PNAS 2008, 105:18970-18975.
38. Goldman MS: Memory without feedback in a neural network.
Neuron 2009, 61:621-634.
39. Lim S, Goldman MS: Noise tolerance of attractor and
feedforward memory models. Neural Comput 2012, 24:332-390.
40. Miri A, Daie K, Arrenberg AB, Baier H, Aksay E, Tank DW: Spatial
gradients and multidimensional dynamics in a neural
integrator circuit. Nat Neurosci 2011, 14:1150-1159.
41. Koulakov A, Raghavachari S, Kepecs A, Lisman J: Model for a
robust neural integrator. Nat Neurosci 2002, 5:775-782.
42. Goldman MS, Levine JH, Major G, Tank DW, Seung HS: Robust
persistent neural activity in a model integrator with multiple
hysteretic dendrites per neuron. Cerebral Cortex 2003,
1:1185-1195.
43. Field GD, Sampath AP, Rieke F: Retinal processing near
absolute threshold: from behavior to mechanism. Annu Rev
Physiol 2005, 67:491-514.
44. Green D, Swets J: Signal Detection Theory and Psychophysics.
Peninsula Pub; 1966.
45. Smith PL, Vickers D: Modeling evidence accumulation with
partial loss in expanded judgment. J Exp Psychol Human
Perception Performance 1989, 15:797.
46. Purcell BA, Heitz RP, Cohen JY, Schall JD, Logan GD, Palmeri TJ:
Neurally constrained modeling of perceptual decision making.
Psychol Rev 2010, 117:1113-1143.
47. Schall JD, Purcell BA, Heitz RP, Logan GD, Palmeri TJ: Neural
mechanisms of saccade target selection: gated accumulator
model of the visual-motor cascade. Eur J Neurosci 2011,
33:1991-2002.
48. Miller P: Power–law neuronal fluctuations in a recurrent
network model of parametric working memory. J Neurophysiol
2006, 95:1099-1114.
49. Cain N, Barreiro A, Shadlen M, Shea-Brown E: A favorable
tradeoff between robustness and performance in sequential
decision tasks. Cosyne Abstracts 2011.
50. Raposo D, Sheppard JP, Schrater PR, Churchland AK:
Multisensory decision-making in rats and humans. J Neurosci
2012, 32:3726–3735.
51. Churchland AK, Kiani R, Shadlen MN: Decision-making with
multiple alternatives. Nat Neurosci 2008, 11:693-702.
52. Roitman JD, Shadlen MN: Response of neurons in the lateral
intraparietal area during a combined visual discrimination
reaction time task. J Neurosci 2002, 22:9475-9489.
53. Shea-Brown E, Gilzenrat M, Cohen JD: Optimization of decision
making in multilayer networks: The role of locus coeruleus.
Neural Comput 2008, 20:2863-2894.
54. Aston-Jones G, Cohen JD: An integrative theory of locus
coeruleus-norepinephrine function: adaptive gain and optimal
performance. Annu Rev Neurosci 2005, 28:403-450.
55. Eckhoff P, Wong-Lin KF, Holmes P: Optimality and robustness of
 a biophysical decision-making model under norepinephrine
modulation. J Neurosci 2009, 29:4301-4311.
The authors include norepinephrine (NE) receptors in a spiking neuron
model of decision making. They show how decision speed and accuracy
depends on receptor densities and NE levels in different cell populations,
identifying ways in which NE can optimize performance for a given
decision making environment and adapt to changes in that environment.
www.sciencedirect.com
 Computational models of decision making Cain and Shea-Brown 1053
The authors also demonstrate how a given NE level can have disparate
effects of circuits with different receptor densities and configurations
56. Usher M, Cohen JD, Servan-Schreiber D, Rajkowsky J, Aston-
Jones G: The role of locus coeruleus in the regulation of
cognitive performance. Science 1999, 283:549-554.
57. Drugowitsch J, Moreno-Bote R, Churchland AK, Shadlen MN,
Pouget A: The cost of accumulating evidence in perceptual
decision making. J Neurosci 2012, 32:3612-3628.
58. Frazier P, Yu A: Sequential hypothesis testing under stochastic
deadlines. In Advances in Neural Information Processing
Systems, vol. 20. Edited by Platt JC, Koller D, Singer Y, Roweis S.
Cambridge, MA: MIT Press; 2008:465-472.
59. Huang Y, Friesen A, Rao R: How prior probability influences
decision making: a unifying probabilistic model. Cosyne
Abstracts 2012.
60. Rao RPN: Decision making under uncertainty: a neural model
based on partially observable markov decision processes.
Front Comput Neurosci 2010, 4.
61. Ditterich J: Evidence for time-variant decision making. Eur J
Neurosci 2006, 24:3628-3641.
62. Wong KF, Huk AC, Shadlen MN, Wang XJ: Neural circuit
dynamics underlying accumulation of time-varying evidence
during perceptual decision making. Front Comput Neurosci
2007, 1:1-11.
63. Dayan P, Yu AJ: Phasic norepinephrine: A neural interrupt
signal for unexpected events. Network: Comput Neural Syst
2006, 17:335-350.
64. Britten KH, Shadlen MN, Newsome WT, Movshon JA: The
analysis of visual motion: a comparison of neuronal and
psychophysical performance. J Neurosci 1992,
12:4745-4765.
65. Britten KH, Shadlen MN, Newsome WT, Movshon JA: Responses
of neurons in macaque MT to stochastic motion signals. Visual
Neurosci 1993, 10:1157-1169.
www.sciencedirect.com
66. Cohen M, Newsome W: Estimates of the contribution of single
 neurons to perception depend on timescale and noise
correlation. J Neurosci 2009, 29:6635-6648.
This paper contributes new evidence for an important experimental
finding, that the time-integral of outputs from MT neurons in the moving
dots task generally approximates the speed and accuracy of decision
making behavior
67. Zohary E, Shadlen MN, Newsome WT: Correlated neuronal
discharge rate and its implication for psychophysical
performance. Nature 1994, 370:140.
68. Mazurek ME, Roitman JD, Ditterich J, Shadlen MN: A role for
neural integrators in perceptual decision making. Cerebral
Cortex 2003, 13:1257-1269.
69. Shadlen M, Britten K, Newsome W, Movshon J: A computational
analysis of the relationship between neuronal and behavioral
responses to visual motion. J Neurosci 1996, 16:1486-1510.
70. Deco G, Rolls ET, Romo R: Stochastic dynamics as a principle
of brain function. Progr Neurobiol 2009, 88:1-16.
71. Smith PL, McKenzie CRL: Diffusive information accumulation
by minimal recurrent neural models of decision making. Neural
Comput 2011, 23.
72. Simen P, Balci F, deSouza L, Cohen JD, Holmes P: A model
of interval timing by neural integration. J Neurosci 2011,
31:9238-9253.
73. Wang X-J: Decision making in recurrent neuronal circuits.
Neuron 2008, 60:215-234.
74. Webb TJ, Rolls ET, Deco G, Feng J: Noise in attractor networks
in the brain produced by graded firing rate representations.
PLoS One 2011, 6:e23630.
75. Doiron B, Litwin-Kumar A, Polk A: Correlated variability
stabilizes persistent activity in networks of spiking neurons.
Abstract 451.23. Soc Neurosci 2011.
76. Boerlin M, Denève S: Spike-based population coding and
working memory.. PLoS Comput Biol 2011, 7:e1001080.
Current Opinion in Neurobiology 2012, 22:1047–1053