In this appendix, we carry out the result shown earlier relating the Bayesian estimation of conditional and joint densities. In presenting these two inference problems, we discussed how they might lead to different solutions despite the fact that they both involve exact techniques (integration and conditioning). In addition, some speculation about the superiority of the direct conditional estimate was made. In the following, we present a specific example to demonstrate this difference and to argue in favor of the conditional estimate p(y|x)c versus the conditioned joint estimate p(y|x)j.
To prove this point, we use a specific example, a simple 2-component
mixture model. Assume the objective is to estimate a conditional
density, p(y|x). This conditional density is a conditioned
2-component 2D Gaussian mixture model with identity covariances. We
try to first estimate this conditional density by finding the joint
density p(x,y) and then conditioning it to get
p(y|x)j. Subsequently, we try to estimate the conditional density
directly to get p(y|x)c without obtaining a joint density in the
process. These are then compared to see if they yield identical
solutions.
Consider a joint density as a two-element 2D Gaussian mixture model
with identity covariance and equal mixing proportions shown in
Figure 11.1. We wish to fit this model to data using
Bayesian integration techniques. The result will not be significant on
its own since this is a trivial learning example. However, we shall
check for inconsistencies between this result and direct conditional
density estimation to prove a general statement about Bayesian
inference. Equation 11.1 depicts the likelihood of
a data point (x,y) given the model and Equation 11.2
depicts a wide Gaussian prior (with very large )
on the
parameters
(m0, n0, m1, n1). As shown earlier, we wish to
optimize this model over a data set
.
This
computation results in a model p(x,y) as in
Equation 11.3.
In Equation 11.3 we are effectively summing over all
the permutations
of the assignments of the N data points to
the 2 different models. For each
,
we select a different
assignment of the i data points. Each point gets assigned to one of
2 Gaussians (one related to
and the other related to
). The summation over the data in each exponential can be
further simplified as in Equation 11.4 and then
analytically integrated. The integrals are summed over all possible
assignments of the data points to one of the two Gaussians (i.e. MNpossibilities or integrals where M=2 models here). Essentially, we
are iterating over all possible permutations where the data points
are assigned to the two different Gaussians all 2N different ways
and estimating the Gaussians accordingly. Evidently this is a slow
process and due to the exponential complexity growth, it can not be
done for real-world applications. Figure 11.2 shows
some data and the Bayesian multivariate mixture model estimate of the
probability density p(x,y). Figure 11.3 shows the
conditional density p(y|x)j.
By solving another integration, we can directly compute the
conditional density p(y|x)c. The conditional density has the form
shown in Equation 11.5. This is just a regular
2-component conditioned mixture of Gaussians model with identity
covariances. Assume that we are using the same prior as before. In
addition, note the presence of the exact same parameters
m0, n0,
m1, n1 which reside in the conditioned parametrization of which can be called
.
The resulting Bayesian integration is depicted in Equation 11.6. Unfortunately, integration can only be completed analytically for the parameters n0 and n1. Thus, the integration over the other 2 parameters is performed using numerical approximation techniques. The inner integral of n0 and n1 causes the exponentially complex assignment permutation seen above and this is compounded with the computation of the integral numerically by a grid approach. This is therefore an even more cumbersome computation than the joint density Bayesian estimate and is only shown here as an example. There exist more efficient numerical integration techniques such as superior quadrature approaches or Monte-Carlo methods however this Bayesian integration approach is typically too intensive for any real-world applications. It should be noted that typically, Bayesian density estimation, Bayesian sampling techniques and Bayesian integration are quite cumbersome except in very special situations.
The same data is thus fitted with the conditional model which produces the conditional distribution p(y|x)c shown in Figure 11.4. Surprisingly, this is quite different from the conditioned joint density. In fact, if we consider a slice of the conditional densities at an arbitrary x value, we obtain the y-distributions shown in Figure 11.5. This indicates that the directly computed conditional model was able to model the bi-modal nature of the data while the conditioned joint density model was not. In fact, p(y|x)c seems like a better choice than p(y|x)j.
The above suggests the following. Consider the case of two Bayesian
statisticians (A and B) who are asked to model a conditional density
(i.e. in a classification task or a regression task) from
data. Statistician A assumes that this conditional density arises from
a joint density. He then estimates this density using full Bayesian
inference. He then conditions this joint density and obtains the final
conditional density he was asked to produce. Statistician B assumes
nothing about the origins of the conditional density and
estimates it directly. He only uses a parametric form for a
conditional density, it is just a function. At the end of the day, the
two have different models even though all the manipulations they
performed where valid equalities (Bayesian inference and conditioning
are exact manipulations). Thus, by a strange by-product of the paths
the two statisticians took, they got two different answers:
.
Typically p(y|x)c seems to be a more robust
estimate and this is probably because no extra assumptions have been
made. In assuming that the original distribution was a joint density
which was being conditioned, statistician A introduced unnecessary
constraints
11.1 from
the space of p(x,y) and these prevent the estimation of a good model
p(y|x). Unless the statisticians have exact knowledge about the
generative model, it is typically more robust to directly estimate a
conditional density than try to recover some semi-arbitrary joint
model and condition it. Figure 11.6
graphically depicts this inconsistency in the Bayesian inference shown
above. Note here that the solutions are found by fully Bayesian
integration and not by approximate MAP or ML methods. Thus, the
discrepancy between p(y|x)j and p(y|x)c can not be blamed on the
fact that MAP and ML methods are just efficient approximations to
exact Bayesian estimation.