Bayes and Big Data: The Consensus Monte Carlo Algorithm

Steven L. Scott
1
, Alexander W. Blocker
1
, Fernando V. Bonassi
1
, Hugh A. Chipman
2
,
Edward I. George
3
, and Robert E. McCulloch
4
1
Google, Inc.
2
Acadia University
3
University of Pennsylvania, The Wharton School
4
University of Chicago, Booth School of Business
October 31, 2013
Abstract
A useful denition of big data is data that is too big to comfortably process on a single
machine, either because of processor, memory, or disk bottlenecks. Graphics processing units
can alleviate the processor bottleneck, but memory or disk bottlenecks can only be eliminated
by splitting data across multiple machines. Communication between large numbers of machines
is expensive (regardless of the amount of data being communicated), so there is a need for
algorithms that perform distributed approximate Bayesian analyses with minimal communica-
tion. Consensus Monte Carlo operates by running a separate Monte Carlo algorithm on each
machine, and then averaging individual Monte Carlo draws across machines. Depending on the
model, the resulting draws can be nearly indistinguishable from the draws that would have been
obtained by running a single machine algorithm for a very long time. Examples of consensus
Monte Carlo are shown for simple models where single-machine solutions are available, for large
single-layer hierarchical models, and for Bayesian additive regression trees (BART).
1 Introduction
This article describes a method of performing approximate Monte Carlo simulation from a Bayesian
posterior distribution based on very large data sets. When the data are too large for a single
processor, the obvious solution is to divide them among multiple processors. There are two basic
methods of computing on the divided data. The rst is to divide the work among multiple cores
on the same chip, either on a multi-core central processing unit (CPU), or on a massively parallel
graphics processing unit (GPU). The multi-core approach can be extremely eective (Suchard,
Wang, Chan, Frelinger, Cron, and West, 2010; Lee, Yao, Giles, Doucet, and Holmes, 2010), but it
has two limitations. The rst is that multi-core computing cannot alleviate bottlenecks related to
memory or disk. The second issue is programming. The multi-threaded code necessary for multi-
core computing can be dicult to write, even for expert programmers, because it is subject to
race conditions that are poorly understood and dicult to debug. GPU programming in particular
1
requires very low level memory management that is dicult to abstract to higher level programming
languages.
The alternative to multi-core computing is multi-machine computing, where data are divided
among many dierent machines. Multi-machine computing provides scalable memory, disk, and
processing power, so it can eliminate bottlenecks in all three traditional computing resources.
However, multi-machine computing comes with a very high cost (in terms of eciency) of com-
municating between machines. The multi-machine and multi-core approaches are complementary.
If a multi-core solution is available it can be embedded in a multi-machine strategy, making each
machine faster.
The primary dierence between the multi-machine and multi-core computing models is com-
munication, which is fast in multi-core systems and slow in multi-machine systems. Dierent com-
munication costs imply that dierent algorithms are necessary for multi-core and multi-machine
environments. On a multi-core system one can eectively parallelize a sequential algorithm by
distributing the work required to implement an expensive step. For example, Suchard et al. (2010)
achieve substantial speed increases by implementing a data augmentation algorithm on a GPU.
Their algorithm is a standard data augmentation with the computationally intensive loop over the
data done in parallel. Section 2 of this article presents a case study demonstrating the ineciency
of that algorithm in a multi-machine environment. It illustrates a fact which is well known to par-
allel computing experts, but often surprising to novices: passing messages among a large number
of machines is expensive, regardless of the size of the messages being passed.
For Monte Carlo algorithms to succeed in a multi-machine environment, they must avoid regular
communication between machines. Consensus Monte Carlo attacks the communication problem by
dividing the data across multiple machines, with each machine independently sampling from the
posterior distribution given its data. Posterior draws from each machine are then combined to form
a consensus, system-wide belief about the model unknowns. The consensus Monte Carlo algorithm
is embarrassingly parallel, as dened by the mapreduce framework (Dean and Ghemawat, 2008),
so it can be run on virtually any system for parallel computing, including Hadoop (White, 2012),
multi-core systems, or networks of workstations (Anderson et al., 1995).
Consensus Monte Carlo is one of several attempts to scale traditional statistical computation.
Zhang, Duchi, and Wainwright (2012), and references therein, describe the literature on averaging
frequentist point estimates for parameters of statistical models. Guha, Kidwell, Hafen, and Cleve-
land (2009) discuss distributed methods of visualizing large data sets. Kleiner, Talwalkar, Sarkar,
and Jordan (2011) extend the bootstrap to distributed data with the bag of little bootstraps. To
the extent that bootstrap samples can be viewed as approximating posterior samples, this work can
be considered an alternative to the work presented here. Returning to Bayesian inference, Huang
and Gelman (2005) proposed a consensus Monte Carlo approach that is similar to ours, but with
a dierent rule for combining posterior draws. For certain classes of models (and certain classes
2
of model summaries), Bayesian inference can be conducted without Monte Carlo through clever
approximations such as INLA (Rue, Martino, and Chopin, 2009), or variational Bayes (e.g. Hinton
and Van Camp, 1993; Ghahramani and Beal, 2001; Jaakkola and Jordan, 2000). These approaches
are eective, but there can be reasons to prefer Monte Carlo. For example, the computational
cost of INLA is exponential in the dimension of the parameter space (or hyperparameter space
in the case of hierarchical models), and the approximate posteriors delivered by variational Bayes
give good approximations to individual marginal distributions, but not to the joint distribution as
a whole. Posterior simulation by Monte Carlo is considerably more expensive than deterministic
approximation algorithms, but it delivers the full posterior distribution. Although the approach
presented here is limited to continuous parameter spaces, Monte Carlo methods can generally be
applied to arbitrarily complex models.
The remainder of the article is structured as follows. Section 2 contains the case study, men-
tioned above, illustrating the high cost of communicating between processes in a multi-machine
environment. Section 3 describes the consensus Monte Carlo algorithm, and then Section 4 pro-
vides a series of examples illustrating consensus Monte Carlo on a variety of simulated and real
data sets. Section 5 is a concluding discussion.
2 Coordinating many machines is expensive
This Section presents timings from a multi-machine MCMC algorithm, with the aim of illustrating
just how expensive multi-machine communication can be. We focus on a single layer hierarchical
logistic regression model
y
ij
Binomial(n
ij
, p
ij
)
logit(p
ij
) = x
T
ij

i
N (, )
| N (0, /)

1
W(I, ),
(1)
where W(I, ) is the Wishart distribution with sum of squares matrix I and scale parameter .
The precise numerical values of and used in the hyperprior are unimportant for this example.
This model was applied to an internet advertising data set, where y
ij
is the number of clicks on
an advertisement with characteristics x
ij
that was shown n
ij
times. The hierarchy corresponds to
dierent internet domains (e.g. espn.com vs. nytimes.com). There are slightly more than 600,000
domains in our data set, each with its own
i
. The vector x
ij
contains an intercept and 7 dummy
variables describing the color, font, and similar characteristics of the advertisement, so that the
dimension of x
ij
is 8. The number of distinct congurations of x
ij
in each domain is small. It
3
(a) (b)
Figure 1: Step times for the naive MCMC algorithm in Section 2 with (a) 500 and (b) 50 machines.
is bounded above by 2
7
= 128, but it is rarely more than a few dozen, and often just a single
observation.
The obvious parallel MCMC algorithm for this model partitions the complete data by domain
across a set of worker machines, and assigns one worker to be the master node responsible for
the prior distribution. The algorithm then alternates between drawing each
i
given current values
of and , and then drawing and given the set of
i
s. During the parallel data augmentation
step, each machine simulates from p(
i
|, , y) for its domains, and adds the draw to a set of
low dimensional sucient statistics (

i

i
,

i

i

T
i
). When a machine nishes drawing
i
for all
of its domains, it sends its sucient statistics to the master node, which accumulates sucient
statistics from dierent machines as they become available. Once all machines have reported, the
master node draws and , broadcasts them to the workers, and the cycle repeats. The algorithm
was implemented in Pregel (Malewicz, Austern, Dehnert, Horn, Leiser, and Czajkowski, 2010), a
system for distributed graph computing written by professional software engineers who are experts
in parallel computation. The statistical computations were done in optimized C++ code.
There are potential issues with the adequacy of this model, the convergence rate of the MCMC
algorithm, and the specic method used to sample the logistic regression coecients. Set all these
aside and focus on the amount of time required to perform an MCMC iteration. Figure 1 plots
the time taken by each step in the algorithm from two separate runs with the same data on 500
and 50 machines. The draws of
i
and (, ) are separate SuperSteps in Pregel, so each panel in
Figure 1 represents 10,000 MCMC iterations, with two SuperSteps per MCMC iterate. The 500-
machine run completed in 2.75 hours, which is very close to 1 iteration per second. The 50-machine
run completed in around 5 hours. Notice that a ten-fold reduction in computing resources only
produced a two-fold increase in compute time, so there is at least a factor of 5 ineciency at play.
Job failures are a fact of life in multi-machine computing and are often discussed as a source
4
of delay. Jobs can fail either because of faulty hardware or because of evictions in favor of higher
priority jobs. To protect against job failures, Pregel periodically saves its state so that it can
recover from a failure without completely restarting. The state-saving iterations are highlighted
with triangles in Figure 1. Both runs lose about the same amount of time saving state. The
iterations containing job failures are highlighted with circles. Each recovery from a failure takes a
few tens of seconds. The 500 machine run experienced more failures, but not enough to materially
factor into the eciency dierence between the two runs.
The source of the ineciency can be seen in the lower of the two black bands in Figure 1. The
bands represent even and odd numbered SuperSteps corresponding to the expensive
i
draws in
the even steps and the cheap (, ) draws in the odd steps. As expected, increasing the number
of machines leads to a decrease in the amount of time spent on the expensive step. The surprising
thing is how much more expensive the cheap steps become. In the 50 machine run, the (, )
draw takes a median of roughly 10
2
= 100 milliseconds. In the 500 machine run it takes roughly
10
2.4
250 milliseconds. Logging in the C++ code shows that the actual time spent drawing
and is less than 1 millisecond, with the rest lost to communication overhead. The ineciency
is not coming from rare, catastrophic machine failures, but the consistently higher communication
cost in iterations where no failures occurred. Ironically, CPU was the limiting resource in the single
machine version of this problem, but on a relative scale we wound up devoting almost no CPU time
to our job!
We wish to emphasize that time is not being lost in what one might think are the obvious places:
an imbalanced workload among the machines, inhomogeneous hardware, an inecient broadcast of
the parameters, or reduction of the sucient statistics, or a long transmission time on the network.
All these possibilities were investigated and found to be performing eciently. The problem is that
we are communicating between machines at all. There is simply an overhead involved in getting a
set of machines ready to send and receive messages, even if those machines are dedicated to your
job. The cost is high enough that, in this case, a 500 machine job could achieve at most 4 iterations
per second, on average, even if the machines had no data to process. The clear lesson is that if
Bayesian posterior simulation is to be distributed across multiple machines, it must be done in a
way that uses minimal communication.
3 Consensus Monte Carlo
The idea behind consensus Monte Carlo is to break the data into groups (called shards), give each
shard to a worker machine which does a full Monte Carlo simulation from a posterior distribution
given its own data, and then combine the posterior simulations from each worker to produce a set of
global draws representing the consensus belief among all the workers. Let y represent the full data,
let y
s
denote shard s, and let denote the model parameters. For models with the appropriate
5
independence structure the system can be written
p(|y)
S

s=1
p(y
s
|)p()
1/S
. (2)
Notice that the prior distribution p() =

s
p()
1/S
is broken into S components to preserve the
total amount of prior information in the system. Equation (2) assumes that batches of observations
are conditionally independent across shards, given parameters, but it allows for arbitrary depen-
dence within the elements of y
s
. Thus, for example, it can be used with hierarchical models with
hyperparameters , as long as data in a single level of the hierarchy are not split across two dierent
machines.
If each worker returned its posterior distribution as a mathematical function, then the worker-
level distributions could be multiplied to form the overall consensus posterior given the complete
data. The diculty is that the workers-level distributions are reported as Monte Carlo draws, so
we need a way of combining the draws.
3.1 Combining draws by weighted averages
Suppose worker s generates draws
s1
, . . . ,
sG
from p(|y
s
) p(y
s
|)p()
1/S
. One way to com-
bine draws across workers is to simply average them. Suppose each worker is assigned a weight,
representable as a matrix W
s
. The consensus posterior for draw g is

g
=
_

s
W
s
_
1

s
W
s

sg
. (3)
When each p(|y
s
) is Gaussian, the joint posterior p(|y) is also Gaussian, and equation (3) can
be made to yield exact draws from p(|y). To see this, suppose S = 2, and that |y
1
N (
1
,
1
),
and |y
2
N (
2
,
2
). Then standard Bayesian calculations for the normal distribution give
p(|y) p(|y
1
)p(|y
2
)
N (| , V )
(4)
where V
1
=
1
1
+
1
2
and = V (
1
1

1
+
1
2

2
). Now let
1
be a draw from N (
1
,
1
) and
let
2
be a draw from N (
2
,
2
). It is easily veried that
V
_

1
1

1
+
1
2

2
_
N ( , V ) . (5)
Taken together, equations (4) and (5) suggest the algorithm in Figure 2. Although the algorithm
is exact only for Gaussian posteriors, there are two reasons to believe that it is broadly useful. The
rst is that under standard regularity conditions for asymptotic theory, posterior distributions tend
6
1. Divide y into shards y
1
, . . . , y
S
.
2. Run S separate Monte Carlo algorithms to sample
sg
p(|y
s
) for g = 1, . . . , G, with each
shard using the fractionated prior p()
1/S
.
3. Combine the draws across shards using weighted averages:
g
= (

s
W
s
)
1
(

s
W
s

sg
).
Figure 2: The consensus Monte Carlo algorithm.
towards a Gaussian limit in large samples (Le Cam and Yang, 2000). The second is that we will see
examples in Section 4 where applying the method to non-Gaussian posteriors (without theoretical
justication) works surprisingly well.
3.2 Choosing Weights
The weight W
s
=
1
s
is optimal (for Gaussian models), where
s
= V ar(|y
s
). An obvious Monte
Carlo estimate of
s
is sample variance of
s1
, . . . ,
sG
. If the dimension of is very large, or if the
model is very simple, one may prefer to weight sub-optimally in order to simplify the computation.
If is high dimensional then it may be preferable to ignore the covariances in
s
, and simply weight
each scalar element of by the reciprocal of its marginal posterior variance.
In many big data problems the analyst is in control of the sharding process, so data can be
sharded by randomly assigning each observation (or cluster of correlated observations) to a shard.
One is tempted to argue that the posterior distributions in each shard will have roughly the same
shape, and so the draws from dierent shards can be combined with equal weighting. However, large
samples on each shard are necessary for the preceding argument to apply. In particular, each worker
must have enough information to estimate all components of . Almost by denition, models that
are complex enough to require very large data sets will have subsets of parameters that cannot be
estimated on every shard. Section 4.3 provides an example where shard-level posterior distributions
can be very dierent, even with IID data, because some observations are more informative than
others. In practice, information-based weighting will usually be necessary.
3.3 Other potential consensus strategies
Strategies other than averaging could be used to form consensus posterior estimates. We focus on
averages because they are simple and stable. For example, one could empirically estimate each
worker-level posterior density using a kernel density estimate and then apply equation (2) directly.
This approach might work well in small models, but as the dimension of increases the kernel
density estimates will become unreliable.
Huang and Gelman (2005) suggest four alternative methods for consensus posteriors, all of which
are variants of either an explicit normal approximation or importance resampling. The consensus
7
Monte Carlo method from Figure 2 is also rooted in normal theory, but it has the potential to
capture non-normal features of the posterior such as fat tails or skewness that would be lost under
an explicit normal approximation. We have not explored importance resampling methods because of
the potential for importance resampling to collapse to a single point in high dimensions, and because
of the potential diculty in evaluating the likelihood p(y
s
|) in complex models, like hierarchical
generalized linear models similar to equation (1). On the other hand, consensus distributions based
on averaging are obviously limited to continuous parameter spaces, while importance resampling
methods could be applied more generally (e.g. to mixtures, or regression models with spike-and-slab
priors).
3.4 More complex models
3.4.1 Nested hierarchical models
If the data have nested structure, where y
ij
f(y|
j
) and
j
p(|), then consensus Monte
Carlo can be applied to the hyperparameters in a straightforward way. If the data are partitioned
so that no group is split across multiple shards, then the model satises equation (2). One can
run the consensus Monte Carlo algorithm, storing the draws of and discarding the draws of
j
.
Combining the draws of
sg
based using empirical weights W
s
estimated from the within-shard
Monte Carlo variance produces a set of draws
1
, . . . ,
G
approximating p(|y).
Conditional on the simulated draws of , sampling
j
p(
j
|y, ) = p(
j
|y
j
, ) is an embar-
rassingly parallel problem. The simulated draws of can be broadcast back to the workers, which
can simulate each
j
in parallel. Each worker machine simulates
jg
given the corresponding
g
,
with no need to communicate with other workers. Because p(|y) the second set of
j
draws
will have marginal distribution p(
j
|y), whereas the rst, discarded set follows p(
j
|y
s
).
This procedure can be recursively applied as needed if the data are nested more than two levels
deep.
3.4.2 Nonparametric regression
In a nonparametric regression model y N
_
f(x),
2
_
the goal is to compute p(f(x)|y) where
the form of f is unknown. Many nonparametric models either have no parameters to average or
else the model parameters are non-numerical constructs like decision trees (see Denison, Mallick,
and Smith, 1998; Chipman, George, and McCulloch, 2010). However, the predictive distributions
from these models are often nearly Gaussian (possibly after conditioning on latent variables), so
consensus Monte Carlo can be used to combine their predictive distributions at particular values
of x. Given training data y and a set of locations x
1
, . . . , x
K
, where predictions are desired, one
can run S separate Monte Carlo simulations to produce draws from p(f(x
k
)|y
s
). At each x
k
,
each workers draws can be weighted by the inverse of the Monte Carlo estimate of the the within-
8
8.8 8.9 9.0 9.1 9.2 9.3 9.4
0
1
2
3
4
5
6
7
parameter
d
e
n
s
i
t
y
Figure 3: Example of small sample bias. The solid line is the posterior distribution of
2
as described in
Section 3.5. The dashed line is the consensus Monte Carlo estimate. The dotted line is the consensus Monte
Carlo estimate with a jackknife bias correction applied.
worker posterior variance s
2
sk
=

G
g=1
(f
sg
(x
k
)

f
s
(x
k
))
2
/(G1), and averaged to form a consensus
posterior distribution.
3.5 The potential for small-sample bias
Oddly enough, small sample bias can play a role when analyzing large data sets. When the data
are divided among many machines, bias that vanishes in the full data may still be present in each
shard. Zhang, Duchi, and Wainwright (2012) discuss the issue of small sample bias in the context
of frequentist point estimation. A similar eect can exist in Bayesian analysis, where the consensus
posterior needs to be shifted because of an accumulation of small sample biases.
Figure 3 shows the eect of small sample bias in a toy simulation where y
i
N (, 1), with
p() 1, where the parameter of interest is
2
. The Figure is based on a simulation of n = 10, 000
observations from y N (3, 1), with 10 equally weighted shards. The solid line in the Figure is
the posterior distribution based on all 10, 000 observations, while the dashed line is the consensus
Monte Carlo estimate.
Not all problems will exhibit meaningful small sample bias. In some cases the shard-level models
will have suciently large sample sizes that small-sample bias is not an issue. In others, models can
be unbiased even in small samples. When small sample bias is a concern jackknife bias correction
can be used to mitigate it through subsampling. The idea behind jackknife bias correction is to
shift the draws from a distribution by a specied amount B determined by subsampling. Suppose
E(|y) = + B/n, so that the posterior mean is a biased estimate of . On each worker machine,
9
take a subsample of size n, where 0 < < 1. Let y
sub
denote the resulting distributed subsample,
and construct consensus Monte Carlo draws from p(|y
sub
). Then E(|y
sub
) = +
B
n
. Simple
algebra gives an estimate of the bias term
B
n
(E(|y
sub
) E(|y))
_

1
_
.
Subtracting the estimated value of B/n from each draw in the original consensus distribution gives
a set of rst order bias corrected draws. The dotted line in Figure 3 shows the eect of the bias
correction in our example.
4 Examples
This section proceeds through a set of examples illustrating the consensus Monte Carlo algorithm.
In each case, the goal is to see how the algorithm compares to a single machine algorithm run on
the same data. Thus the examples will have to be simple enough (or else the data small enough)
for a single machine run to be possible.
4.1 Binomial data with a beta prior
Figure 4(a) shows three Monte Carlo approximations to the posterior distribution of a binomial
success probability. The data are 1000 Bernoulli outcomes with a single success, and the assumed
prior is uniform. Two of the distributions in Figure 4 are made by directly drawing from the
Be(2, 1001) distribution. The distribution marked Consensus was constructed by distributing
the data across 100 independent Monte Carlo samplers, each with a Be(.01, .01) prior distribution.
One of the samplers is assigned the shard of data containing 9 failures and the single success. The
remaining 99 samplers each observe 10 failures. The posterior draws were then combined using
an unweighted average. This example is interesting because even though the posterior is highly
non-Gaussian, consensus Monte Carlo does an eective job capturing the distribution. Figure 4(b)
highlights that the approximation is imperfect, in that it slightly under-weights the tails.
This example also illustrates that the procedure can be sensitive to the scale of the analysis.
The shard-level prior Be(.01, .01) is not the same as the full Be(1, 1) prior raised to the .01 power.
If the 100 workers had each been given the uniform prior then the estimated distribution would
be seriously incorrect, as shown by Figure 5(a). The prior is very important because only a
single success was observed. Giving each worker an additional prior success corresponding to
the uniform prior adds a substantial amount of prior information to the system. The inuence of
the prior, and thus the need for prior adjustment, becomes less dramatic when the data are more
informative.
Figure 5(b) shows a dierent simulation where the workers are given dierent amounts of work
10
0.000 0.002 0.004 0.006 0.008 0.010 0.012
0
1
0
0
2
0
0
3
0
0
N = 10000 Bandwidth = 0.0001827
D
e
n
s
i
t
y
Single Machine
Consensus
Another Single Machine
0.000 0.002 0.004 0.006 0.008 0.010 0.012
0
.
0
0
0
0
.
0
0
2
0
.
0
0
4
0
.
0
0
6
0
.
0
0
8
0
.
0
1
0
Single Machine
C
o
n
s
e
n
s
u
s
(a) (b)
Figure 4: (a) Posterior draws from binomial data. (b) A qq plot showing that the tails of the consensus
Monte Carlo distribution in panel (a) are slightly too light.
to do. There were 5 workers each assigned 100, 20, 20, 70, and 500 binomial observations. The data
were simulated from the binomial distribution with success probability .01. The workers were each
given a Be(1/5, 1/5) prior to match the Be(1, 1) uniform prior used in the single machine analysis.
This distribution is slightly less skewed than the one in Figure 4, because the simulation produced
more than one success. The distribution is still noticeably non-Gaussian, and the consensus Monte
Carlo estimate is very similar to the two single machine analyses.
4.2 Gaussian Data
The next test case is a multivariate normal model with unknown mean and variance. The test case
has = (1, 2, 3, 4, 5) and
=
_
_
_
_
_
_
_
_
1.00 0.99 0.98 0.00 0.70
0.99 1.00 0.97 0.00 0.75
0.98 0.97 1.00 0.00 0.60
0.00 0.00 0.00 1.00 0.00
0.70 0.75 0.60 0.00 1.00
_
_
_
_
_
_
_
_
.
The model includes three deviates that are highly correlated with one another, one that is indepen-
dent of everything else, and one that is anticorrelated with the rst three. We consider a problem
with 100 workers and examine how consensus Monte Carlo handles draws of . We focus on
because the marginal posterior of is multivariate T, which is suciently close to normal for the
11
0.00 0.02 0.04 0.06 0.08 0.10
0
1
0
0
2
0
0
3
0
0
N = 1000 Bandwidth = 0.0002871
D
e
n
s
i
t
y
Single Machine
Consensus
Another Single Machine
0.000 0.005 0.010 0.015 0.020 0.025 0.030
0
2
0
4
0
6
0
8
0
1
0
0
1
2
0
N = 10000 Bandwidth = 0.0004677
D
e
n
s
i
t
y
Single Machine
Consensus
Another Single Machine
(a) (b)
Figure 5: The consensus Monte Carlo distribution (a) performs badly when each worker receives a uniform
prior, and (b) performs well with imbalanced shards, when properly weighted.
consensus draws to be nearly exact.
Figure 6 shows posterior draws of based on simulated data with 50 observations per worker.
The scale is dierent in each of the panels (which are arrayed as the elements of ) to maximally
highlight dierences in the consensus and single-machine draws. When viewed on the same scale,
the dierences between parameter values are far more important than the dierences between
methods. The consensus Monte Carlo estimates (shown as green dashed lines) of the posterior
appear to have a bias away from zero when compared to the single machine algorithm (the solid
lines).
The dotted blue lines are a jackknife bias corrected density estimate based on a 20% subsample.
With 50 observations per worker, a 20% subsample is 10 observations. When paired with corrected
prior distribution (obtained by dividing both the prior sum of squares and prior sample size by
S), a much smaller sub-sample would lead to an improper posterior. The bias correction generally
seems to help in Figure 6(a), though for the variance term in element 4 it provides an adjustment
were none is necessary.
Interestingly, when we double the sample size to 100 in Figure 6(b) the unadjusted consensus
Monte Carlo estimates match the single machine algorithm quite closely, while the jackknife bias
adjustment winds up introducing a bias towards zero similar to the one it removed in Figure 6(a).
We tried one more run with 1000 observations per worker, which we chose not to display because all
distributions matched suciently closely that it was dicult to visually distinguish between them.
12
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
parameter
d
e
n
s
ity
(a) (b)
Figure 6: (a) Posterior distribution of based on 100 workers, with 50 observations per worker. Red
(solid) line is the single machine algorithm. Green (dashed) line is the consensus Monte Carlo estimate.
Blue (dotted) line is the bias corrected consensus estimate. (b) Same, but with 100 observations per worker.
With 1000 observations per worker all plots overlap essentially perfectly.
4.3 Logistic regression
Logistic regression is a widely used model for internet data because of the interest in whether a
particular stimulus results in a user action such as a click, mouse-over, or similar observable event.
Table 1(a) is a set of simulated data based on a hypothetical set of binary predictors that occur
independently with the frequencies shown in Table 1(b). The rst variable is an intercept, and the
last is a rarely occurring variable that is highly predictive of an event when it occurs.
Figure 7(a) shows the posterior distribution of the coecients in a logistic regression of y on x
1
through x
5
. The single machine overall MCMC estimate is compared to consensus Monte Carlo
estimates under three dierent weighting schemes. The matrix scheme weights each workers
draws using the inverse of the sample variance matrix for draws from that worker. The scalar
weighting scheme weights each individual component of by the inverse of its scalar sample vari-
ance, ignoring covariances. The equal scheme simply averages all draws with equal weights.
Despite the fact that all 100 workers faced identical sampling distributions, the equal weighting
scheme fails badly because the workers are only equal a priori. They dont see the same data.
Out of 10,000 total observations, there are only 104 in which x
5
is active, and only 71 of those
produce an event. Thus, many workers (roughly 1/3) have no information with which to measure
the eect of x
5
, and roughly another third have only a single observation, so a small subset of
workers has all the information about this coecient. The dierent subsets of workers can be seen
13
y n x
1
x
2
x
3
x
4
x
5
266 2755 1 0 0 1 0
116 2753 1 0 0 0 0
34 1186 1 0 1 0 0
190 717 1 1 0 1 0
61 1173 1 0 1 1 0
37 305 1 1 1 0 0
68 301 1 1 1 1 0
119 706 1 1 0 0 0
18 32 1 0 0 0 1
13 17 1 0 1 1 1
18 24 1 0 0 1 1
8 10 1 1 0 1 1
2 2 1 1 1 0 1
7 13 1 0 1 0 1
2 2 1 1 1 1 1
3 4 1 1 0 0 1
x
1
x
2
x
3
x
4
x
5
frequency 1 .2 .3 .5 .01
coecient -3 1.2 -.5 .8 3
(a) (b)
Table 1: (a) Data for the logistic regression in Figure 7. (b) The probability that each variable is active,
and the true logistic regression coecients used in the simulation.
in Figure 7(b). Those with no information about
5
sample from the prior. Those that only get
to see a single observation sample from one of the the skewed distributions, depending on whether
the single observation corresponded to a success or a failure. The subset of informed workers able
to see multiple observations sample from the nearly normal distributions centered around 3. The
empirical weighting schemes (matrix and scalar) are aware of the information asymmetry and
can use it to place more weight on workers that have more information.
The information asymmetry is the main reason the equal weighting scheme misses so badly
in Figure 7(a). The Figure also shows substantial agreement between the single-machine overall
MCMC algorithm and the matrix and scalar weighting schemes. Both schemes do a good job
capturing not only the location and spread of the marginal distribution of the coecients, but also
the correlations between coecients. The scalar scheme tends to be over-dispersed relative to
the matrix scheme, but not terribly so.
Figure 8 shows how the various schemes perform as the sample size grows to 1,000 and 10,000
observations per worker. Both the matrix and scalar weighting methods do about equally well
combining the worker draws. The equal weighting scheme is not shown in these plots. It still
missed badly with 1000 observations per worker, but it performs about as well as the other methods
with 10,000 observations per worker. With so many observations, the dierences between pairwise
marginal distributions are vanishingly small when seen on a common scale.
14
parameter
d
e
n
s
it
y
x1
2
4
6
1
2
3
4 1
2
3
4
0.5
1
1.5
2
2
.5

3
7

x2
2
1
2
0.5
1
0.2
0.4
0
.6

1

3
x3
2
4
6
1
2
3 1
2
3
4
0.5
1
1.5
2
2
.5

3

x4
0
.5

0.2
0.4
0.2
0.4
0.05
0.1
0
.1
5

0
.2

0
.6
5

x5
x
1

4
0
2
4
6
parameter
d
e
n
s
it
y
1

2

1
2
3
0
.5

1
1
.5

2

0.2
0.4
0
.6

0
.8

1

2
4
6

1
2
3
4

5

6

0.5
1
1
.5

2
2
.5

3
1
0

0.5
1
1.5
2
2
.5

3

5

6
.5

0
.5

1 0.2
0.4
0
.6

0
.8

0.2
0.4
0
.6

0
.8

0.1
0.2
0
.3

0
.4

0
.5

1
.1

x
2
parameter
d
e
n
s
it
y
1
2
1
2
0.5
1
1.5
2
0.2
0.4
0.6
0
.8

1

2
.6

1
0.2 0.1
0.2
0.05
0.1
0
.1
5

0
.2
x
3

4
0
2
4
6
parameter
d
e
n
s
it
y
0
.5

1

0.2
0.4
0
.6

1

0.2
0.4
0
.6

0
.8

0.05
0.1
0
.1
5

0
.2

0.55
0.65
x
4
0
.0
5

4 0 2 4 6
1
1

0
.1

1

0
.1

0
.0
5

4 0 2 4 6
1

0
.0
5

parameter
d
e
n
s
it
y
x
5
overall matrix scalar equal
40 20 0 20 40
0
.
0
0
0
.
0
5
0
.
1
0
0
.
1
5
0
.
2
0
0
.
2
5
parameter
d
e
n
s
i
t
y
(a) (b)
Figure 7: Logistic regression with 100 observations per worker. Panel (a) is a pairs plot comparing the
single machine MCMC estimate of the posterior distribution with the consensus Monte Carlo estimate under
three dierent weighting schemes. Plots on and above the diagonal show the distributions on a scale just large
enough to t the plots in each dimension. Plots below the diagonal are on a common scale. The diagonal
shows the marginal distribution of each coecient. Panel (b) shows the worker-level marginal distributions
for
5
.
4.4 Hierarchical models
We now consider tting a hierarchical model to a large distributed data set of internet advertising
data. The model is the hierarchical Poisson regression described in equation (6), where y
ij
is the
number of times advertisement i from advertiser j was clicked, E
ij
is the number of times it was
shown, and x
ij
is a small set of predictor variables, including dummy variables describing the type
of ad format the ad was shown in (e.g. part of a vertical vs. horizontal conguration, along the
top vs. along the side, etc.), and a continuous quality score assigned by another model.
y
ij
Po(E
ij

ij
)
log
ij
=
T
j
x
ij

j
N (, )
N (0, /)

1
W(I, )
(6)
15
parameter
d
e
n
s
ity
x1
20
40
6
0

8
0

1
0
0

2
2
0

10
20
30
40
50
6
0

7
0

120
10
20
30
40
5
0

6
0

7
0

8
0

x2
10
10 10
2
0

30
4
0

5
0

60
9
0

170 180
10
20
30
40 5
0

6
0
7
0

100
110
10
20
30
40
5
0

6
0

7
0

110
x3
20
40
60
80
1
0
0

10
20
30
40
5
0

60
7
0

8
0

160
10
20
30
40
5
0

6
0

7
0

150
x4
5
5

10
1
5

2
0

2
5

3
0

5
0

80
5
10
15
2
0

2
5

3
0
3
5

60
5
10
15
2
0

2
5

3
0

35
50
x5
x
1

1
1
2
3
parameter
d
e
n
s
ity
20
40 60
80
1
0
0

200
10 20
30
4
0

5
0

6
0

70
1
0
0

10
20 30
40 5
0

6
0

7
0

90
10 10
10
20
30
40
5
0
6
0
140
10
20
30
40
5
0
6
0

7
0

100
10
20
30
40 50
6
0
7
0

120
5
5
10
15
2
0

2
5

70
5
10
1
5

20 2
5

30
3
5

5
0

5
10
1
5
20
2
5

30
3
5

4
0

x
2
parameter
d
e
n
s
ity
10
20
30
40
5
0

6
0

120
10
20
30
40
5
0

6
0

7
0

1
1
0

10
20
30
40
5
0

60
7
0

1
1
0

5
10
15
20
2
5

50
5
10
1
5
20
2
5
3
0

40
2
4
6
8 1
0

1
2
1
4

1
6

1
8
2
0

30
3
6

x
3

1
1
2
3
parameter
d
e
n
s
ity
5
10 15
20
2
5

3
0

8
0

5
10
15 2
0

2
5
3
0

40
5
10
15
20
25
3
0

x
4
3 1 1 2 3 3 1 1 2 3
parameter
d
e
n
s
ity
x
5
overall matrix scalar
parameter
d
e
n
s
ity
x1
200
400
6
0
0

2
2
0
0

100
200
3
0
0

4
0
0

5
0
0
1
1
0
0

100
200
300
4
0
0

5
0
0

1
5
0
0

x2
200
400
6
0
0
8
0
0

100
200
3
0
0

4
0
0

5
0
0

6
0
0

1200
100
200
300
400
5
0
0

6
0
0
1400
x3
200
400
600
8
0
0

200
400
600
800
1
0
0
0

200
400
600
800
1
0
0
0

x4
50
100
1
5
0

2
0
0

550
750
50
100
150
2
0
0

2
5
0

50
100
1
5
0

2
0
0
2
5
0

500
x5
x
1

1
1
2
3
parameter
d
e
n
s
ity
200
400
6
0
0

8
0
0

100
200
300
4
0
0

5
0
0

1000
1
2
0
0

100
200
3
0
0

4
0
0

5
0
0

1100
200
400
6
0
0

1400
100
200
300
4
0
0

5
0
0

1200
100
200
300 4
0
0

5
0
0

6
0
0

1
0
0
0

50
100
1
5
0

2
0
0

500
50
100
150
2
0
0

2
5
0

50
100
1
5
0
2
0
0

2
5
0

5
0
0

x
2
parameter
d
e
n
s
ity
200
400
6
0
0

8
0
0

100
200
300
4
0
0

5
0
0

900
1200
100
200
3
0
0

400
600
900
50
100
1
5
0

2
0
0

350
50
100
1
5
0

2
0
0

4
0
0

50
100
1
5
0

2
0
0

250
500 x
3

1
1
2
3
parameter
d
e
n
s
ity
50
100
1
5
0
2
0
0

500
50
100
150
2
0
0

2
5
0

5
5
0

50
100
150
2
0
0

2
5
0

5
5
0

x
4
3 1 1 2 3 3 1 1 2 3
parameter
d
e
n
s
ity
x
5
overall matrix scalar
(a) (b)
Figure 8: Logistic regression with (a) 1,000 and (b) 10,000 observations per worker. Compare to Figure 7(a).
The equal weighting scheme is not shown here, to better focus on matrix and scalar schemes.
The data consist of nearly 24 million observations regarding the performance of individual
advertising creatives nested within around 11,000 advertisers using a particular Google advertising
product. The data were divided into 867 shards, so that no shard contained more than 50,000
observations. A few shards contained between 10,000 and 20,000 observations, with a median
shard size of about 27,000 observations. Most shards were close to the median.
The model from equation (6) was run on a single machine for 10,000 MCMC iterations using the
MCMCglmm package from CRAN (Hadeld, 2010). Figure 9 shows the posterior draws of based
on the rst 5 shards of data under a single machine run and the consensus Monte Carlo algorithm
with 5 workers. We used only 5 of the 867 shards because of the long compute time for the single
machine run. The shards were combined by averaging draws based on the scalar precisions of each
element of . There is general agreement between the single machine and consensus algorithms.
The largest disagreement is on AdFormat 6, which only one of the 5 shards had any information
about. There was one shard that exhibited slow mixing on the AdFormat 5 coecient, leading to
a somewhat overdispersed consensus distribution. That could presumably be improved through a
longer run time or by using a more ecient Monte Carlo algorithm on the worker machine. There
were only two shards with information about AdFormat 4, but its single machine and consensus
estimates agree closely. The posterior correlations between elements of are not particularly
strong, but to the extent that non-spherical errors are present (e.g. between AdFormat 4 and the
Intercept), they appear to be well captured by consensus Monte Carlo.
Figure 10 shows what happens to the computing time as the number of shards increases. The
16
parameter
d
e
n
s
i
t
y
Intercept
0.05
0.1
0.05
0.1
Quality
0.1
0
.2

0.1
0.2
AF 1
0.05
0.05
AF 2
0
.0
5

0
.1

0.05
0.1
AF 3
0.02
0.02
AF 4
0
.0
2

0.02
AF 5
0.01
AF 6
I
n
t
e
r
c
e
p
t

1
0
5
parameter
d
e
n
s
i
t
y
0.05
0.05
0.1
0.02
0.02
0
.0
5

0.02
0.04
0.01
0
.0
1
0.01
0
.0
2

Q
u
a
l
i
t
y
parameter
d
e
n
s
i
t
y
0.05
0.05
0.05
0.05
0
.0
2

0.02
0.02
0
.0
1

A
F

1

1
0
5
parameter
d
e
n
s
i
t
y
0.02
0.02
0.01 0.01

0
.
0
1

0.01
A
F

2
parameter
d
e
n
s
i
t
y
0.02
0.01
0.02
0
.0
1

A
F

3

1
0
5
parameter
d
e
n
s
i
t
y
0.005
A
F

4
parameter
d
e
n
s
i
t
y
0
.0
0
2

A
F

5
10 5

1
0
5
10 5 10 5 10 5
parameter
d
e
n
s
i
t
y
A
F

6
overall consensus
parameter
d
e
n
s
ity
Intercept
parameter
d
e
n
s
ity
Quality
parameter
d
e
n
s
ity
AdFormat 1
parameter
d
e
n
s
ity
AdFormat 2
parameter
d
e
n
s
ity
AdFormat 3
parameter
d
e
n
s
ity
AdFormat 4
parameter
d
e
n
s
ity
AdFormat 5
parameter
d
e
n
s
ity
AdFormat 6
1 2 3 4 5
(a) (b)
Figure 9: Posterior draws of based on the rst 5 shards described in Section 4.4. (a) Posterior draws
from the single-machine and consensus Monte Carlo algorithms. (b) Draws from the ve worker machines.
single machine algorithm scales linearly up to about 20 shards, at which point it encounters a
resource limit and begins slowing down. We stopped experimenting with the single machine run
after 30 shards, because resource bottlenecks caused the machine to stop functioning, and we could
not get the run to nish. With consensus Monte Carlo the job completes within about 20-30
minutes, regardless of the number of shards. These times are with o-the-shelf R code. Optimized
code would improve both the single machine behavior and the run time on each worker machine.
In this application, the estimands of interest require the advertiser specic coecients
j
. As
mentioned in Section 3.4.1, these can be obtained with independent MCMC runs on each adver-
tisers data, given the draws of and .
4.5 Nonparametric regression
We consider nonparametric regression models of the form
y
i
N
_
f(x
i
),
2
_
(7)
17
Figure 10: Time required to complete 10,000 MCMC draws with dierent numbers of shards under the single
machine and consensus Monte Carlo algorithm.
independently across observations. Such models often lack interpretable sets of model parameters,
so Bayesian inference focuses on the posterior distribution p(f(x)|y) at a particular value of x. As
mentioned in Section 3.4.2, Consensus Monte Carlo can be applied to nonparametric regression
models by tting a separate model on each worker machine and averaging the draws from the
worker-level predictive distributions.
As an example, consider BART (Chipman, George, and McCulloch, 2010), a tree-based algo-
rithm for nonparametric regression, where
f(x) =
T

t=1
f
t
(x) (8)
and each f
t
(x) is a decision tree. Each interior node j of tree t compares a specic element
tj
of
x to a specic cutpoint
tj
. Starting at the root of the tree, if node j has children and x

tj

tj
then x falls to the left child, otherwise it falls to the right. If node j is a leaf then it contains a
mean parameter M
tj
giving the contribution of tree t to f
t
(x). BART is an additive model that
combines contributions from many trees.
The prior distribution for BART recommended by Chipman et al. (2010) consists of a prior over
the topology of each tree, a prior over the parameters M
jt
, and a prior on the residual variance

2
. The prior probability that a node at depth d splits into children at depth (d +1) is a/(1 +d)
b
.
If there is a split, a variable
tj
is chosen uniformly from the set of available variables, and a
cutpoint
tj
uniformly from the set of available cutpoints. If there is no split, then the node is a
leaf containing mean parameter M
tj
with prior distribution M
tj
N
_
0,
2
_
, independently across
t and j. A conditionally conjugate inverse gamma prior is assigned to
2
.
We tested the consensus Monte Carlo algorithm with BART using simulated data from Fried-
18
5 10 15 20
5
1
0
1
5
2
0
bart mean
c
o
n
s
e
n
s
u
s
b
a
r
t, n
o
p
r
io
r
a
d
ju
s
tm
e
n
t
5 10 15 20
5
1
0
1
5
2
0
bart 5%
c
o
n
s
e
n
s
u
s
b
a
r
t, n
o
p
r
io
r
a
d
ju
s
tm
e
n
t
5 10 15 20 25
1
0
1
5
2
0
bart 95%
c
o
n
s
e
n
s
u
s
b
a
r
t, n
o
p
r
io
r
a
d
ju
s
tm
e
n
t
5 10 15 20
1
0
1
5
2
0
bart mean
c
o
n
s
e
n
s
u
s
b
a
r
t, p
r
io
r
a
d
ju
s
tm
e
n
t
5 10 15 20

5
0
5
1
0
bart 5%
c
o
n
s
e
n
s
u
s
b
a
r
t, p
r
io
r
a
d
ju
s
tm
e
n
t
5 10 15 20 25
2
0
2
5
3
0
3
5
bart 95%
c
o
n
s
e
n
s
u
s
b
a
r
t, p
r
io
r
a
d
ju
s
tm
e
n
t
Figure 11: Comparing consensus BART with the single machine algorithm. The three columns compare
posterior means (left) and posterior 5% (center) and 95% (right) percentiles from the single-machine and
consensus posterior distributions. In the top row the 30 workers were run with the same prior as the single
machine. In the bottom row the prior was raised to the 1/30 power.
mans test function (Friedman, 1991)
f(x) = 10 sin(x
1
x
2
) + 20(x
3
0.5)
2
+ 10x
4
+ 5x
5
+ (9)
where N
_
0, 3
2
_
is an independent error term. Our full data set contained 20,000 observa-
tions where each component of x
i
was independently simulated from U(0, 1). We also simulated
x
6
, . . . , x
10
as additional superuous predictors. We chose 20,000 observations because that is close
to the largest data set that the single-machine BART algorithm could handle in a reasonable time
period. The consensus Monte Carlo algorithm was implemented using the parallel package from R
on a 30 processor machine.
Figure 11 summarizes the results of two model runs on the same set of simulated data. In the
top row, the single machine algorithm and the worker machines in consensus Monte Carlo each used
the recommended prior from Chipman et al. (2010). In the bottom row the prior was raised to the
1/30 power to account for the fact that there are 30 workers. With no prior adjustment the single
19
machine and consensus algorithms essentially agree on both the posterior mean and the upper
and lower posterior quantiles. When the prior is adjusted, the consensus Monte Carlo posterior
becomes overdispersed. The prior distribution plays a signicant role in determining the size of
each tree in the BART algorithm. Weakening the prior results in much larger trees, and generally
more diuse posterior distributions. Averaging over shards mitigates the overdispersion, but not
enough to match the single machine algorithm. Alternative strategies for prior adjustment, such
as only modifying the prior on M
tj
and
2
, might still prove to be eective, and could potentially
be necessary in other applications.
Interestingly, the consensus Monte Carlo algorithm actually ts the data slightly better than
the single machine algorithm. The correlation between the posterior means and the true function
values for the single machine run was .992. The same correlation for consensus Monte Carlo was
.995. The reason is that the Monte Carlo algorithm used to t BART tends to have trouble mixing
as sample sizes grow large. The consensus Monte Carlo posterior is based on 30 rapidly mixing
workers, while the single machine algorithm is based on one slow mixing worker.
5 Discussion
The idea of distributing a Bayesian calculation to a group of independent workers is a natural one,
justied by equation (2). This article has shown that there are models, some with non-Gaussian
posteriors, for which the draws from workers can be combined using simple weighted averages.
Consensus Monte Carlo scales to very large numbers of machines in ways that multi-core algo-
rithms cannot. It can be implemented with existing code, whether single or multi-threaded, and it
is agnostic to the specic algorithm used to generate the worker-level draws. It is similarly agnostic
to the method used to parallelize the computation. It can be used on a cluster of computers, a
single multi-core or multi-processor computer, or an arbitrary collection of machines that need not
be linked by a high speed network.
A fundamental weakness of the algorithm is its limitation to unknowns of xed dimension in
continuous parameter spaces. For example, label switching and changing numbers of dimensions
would make it dicult to use consensus Monte Carlo to cluster observations based on innite
Dirichlet process mixtures. Similar issues make consensus Monte Carlo ill suited to model averaging
using spike and slab priors.
There is a long list of open questions that needs to be addressed on consensus Monte Carlo.
We need to better understand how the algorithm behaves as posterior distributions move away
from Gaussianity. To the extent that averaging fails in non-Gaussian models other methods of
consensus should be explored. Improvements to the algorithm will probably involve a small num-
ber of between-machine communications, a handful of which would not fundamentally change the
character of the method. Consensus Monte Carlo is likely to be useful with models that violate
equation (2), such as hierarchical models with crossed random eects, or Gaussian processes with
20
non-trivial covariances functions, but these directions need to be explored.
The examples in Section 4 provide useful but sometimes contradictory lessons for practitioners.
Consensus Monte Carlo was found to perform well on a variety of models, but subtle adjustments to
the prior were sometimes necessary to achieve good results. In Section 4.1 adjusting the prior was
critical to the methods success, while in Section 4.5 it was detrimental. Until a more solid theory
can be established, the best advice we can give practitioners is to try consensus Monte Carlo on a
simulated data set to determine conditions under which it performs well in a particular problem.
References
Anderson, T. E., Culler, D. E., and Patterson, D. (1995). A case for now (networks of workstations).
Micro, IEEE 15, 5464.
Chipman, H. A., George, E. I., and McCulloch, R. E. (2010). Bart: Bayesian additive regression
trees. The Annals of Applied Statistics 4, 266298.
Dean, J. and Ghemawat, S. (2008). Mapreduce: simplied data processing on large clusters.
Communications of the ACM 51, 107113.
Denison, D. G. T., Mallick, B. K., and Smith, A. F. M. (1998). Automatic Bayesian curve tting.
Journal of the Royal Statistical Society, Series B, Methodological 60, 333350.
Friedman, J. H. (1991). Multivariate adaptive regression splines. The Annals of Statistics 19, 167.
Ghahramani, Z. and Beal, M. J. (2001). Propagation algorithms for variational Bayesian learning.
Advances in neural information processing systems 507513.
Guha, S., Kidwell, P., Hafen, R., and Cleveland, W. S. (2009). Visualization databases for the
analysis of large complex datasets. In International Conference on Articial Intelligence and
Statistics, 193200.
Hadeld, J. D. (2010). Mcmc methods for multi-response generalized linear mixed models: The
MCMCglmm R package. Journal of Statistical Software 33, 122.
Hinton, G. E. and Van Camp, D. (1993). Keeping the neural networks simple by minimizing the
description length of the weights. In Proceedings of the sixth annual conference on Computational
learning theory, 513. ACM.
Huang, Z. and Gelman, A. (2005). Sampling for Bayesian computation with large datasets. Tech.
rep., Columbia University Department of Statistics.
Jaakkola, T. S. and Jordan, M. I. (2000). Bayesian parameter estimation via variational methods.
Statistics and Computing 10, 2537.
21
Kleiner, A., Talwalkar, A., Sarkar, P., and Jordan, M. I. (2011). A scalable bootstrap for massive
data. arXiv preprint arXiv:1112.5016 .
Le Cam, L. M. and Yang, G. L. (2000). Asymptotics in Statistics: Some Basic Concepts. Springer-
Verlag.
Lee, A., Yao, C., Giles, M. B., Doucet, A., and Holmes, C. C. (2010). On the utility of graphics
cards to perform massively parallel simulation of advanced Monte Carlo methods. Journal of
Computational and Graphical Statistics 19, 769789.
Malewicz, G., Austern, Matthew H. Bik, A. J. C., Dehnert, J. C., Horn, I., Leiser, N., and Cza-
jkowski, G. (2010). Pregel: A system for large-scale graph processing. In SIGMOD10, 135145.
Rue, H., Martino, S., and Chopin, N. (2009). Approximate bayesian inference for latent gaussian
models by using integrated nested laplace approximations. Journal of the royal statistical society:
Series b (statistical methodology) 71, 319392.
Suchard, M. A., Wang, Q., Chan, C., Frelinger, J., Cron, A., and West, M. (2010). Understanding
GPU programming for statistical computation: Studies in massively parallel massive mixtures.
Journal of Computational and Graphical Statistics 19, 419438.
White, T. (2012). Hadoop: the denitive guide. OReilly.
Zhang, Y., Duchi, J. C., and Wainwright, M. J. (2012). Communication-ecient algorithms for
statistical optimization. In Decision and Control (CDC), 2012 IEEE 51st Annual Conference
on, 67926792. IEEE.
22