Auxiliary Variational MCMC


Abstract

Auxiliary Variational MCMC = (1) + (2)

  • (1) MCMC kernels
  • (2) Variational Inference


Exploits LOW dimensional structure in the target distribution,

in order to learn a MORE EFFICIENT MCMC sampler


1. Introduction

Naive pairing of (1) & (2) :

  • use “variational approximation \(q\)” as a “proposal distribution” in a M-H sampler


This paper suggests alternative approach, inspired by..

  • 1) BBVI (Black-Box Variational Inference)
  • 2) auxiliary variable MCMC ( ex. HMC )


Contribution

  • 1) general framework for VI + MCMC
  • 2) use auxiliary variational method to capture latent low-dim structure in target distn
  • 3) extension of M-H algorithm to continuous mixture proposals
  • 4) introduce Auxiliary Variational Sampler (AVS)
    • use flexible distns, parameterized by NN


2. Auxiliary Variational MCMC

Key idea : exploit structure present in the target distn \(p(x)\)

  • by fitting parameterized variational approximation in the AUGMENTED space

    \(\rightarrow\) allows to learn “low-dimensional” structure

Introduce auxiliary variable method & how to combine with proposal distns


2-1. Mixture Proposal MCMC

Notation :

  • \(\tilde{q}\) : proposal distn
  • \(q\) : variational distn


for valid MCMC, need ERGODIC Markov chain, whose stationary distn = \(p(x)\) ( target distn )

\(\rightarrow\) introduce M-H like algorithm, with a specially chosen form of proposal distn

( This proposal can be combined with auxiliary variational method !)

(1) Mixture proposal distribution

  • \(\tilde{q}\left(x^{\prime} \mid x\right)=\int \tilde{q}\left(x^{\prime} \mid a\right) \tilde{q}(a \mid x) d a\).

  • sampling step )

    1. Sample \(a\) from \(\tilde{q}(a \mid x)\)

    2. Sample \(x^{\prime}\) from \(\tilde{q}\left(x^{\prime} \mid a\right)\)

    3. Accept the candidate sample \(x^{\prime}\) with probability

      \(\min \left\{1, \frac{\tilde{q}(x \mid a) \tilde{q}\left(a \mid x^{\prime}\right) p\left(x^{\prime}\right)}{\tilde{q}\left(x^{\prime} \mid a\right) \tilde{q}(a \mid x) p(x)}\right\}\).

      otherwise reject \(x^{\prime}\) and define the new sample as a copy of the current \(x,\) namely \(x^{\prime}=x\).

  • not same as simply performing in joint \((x,a)\)

    ( it is just a sampler in \(x\) alone! )


This mixture proposal can be combined with auxiliary variational method!

2-2. The Auxiliary Variational Method

Auxiliary Variational Method

  • to create more expressive families

  • minimize KL-divergence (X) minimize divergence in an augmented space \((x,a)\), where \(a\) is auxiliary variable (O)

  • How?

    • first, define a joint \(q_{\phi}(x,a)\) ( in augmented space ) & joint distn \(p(x, a)=p_{\theta}(a \mid x) p(x)\).
    • marginalize \(p(x, a)\) over \(a\) \(\rightarrow\) recovers \(p(x)\)
  • objective :

    \[\begin{aligned}\left(\phi^{*}, \theta^{*}\right)&=\underset{\phi, \theta}{\arg \min } \mathrm{KL}\left(q_{\phi}(x, a) \mid \mid p(x,a)\right)\\ &= \underset{\phi, \theta}{\arg \min } \mathrm{KL}\left(q_{\phi}(x, a) \mid \mid p(x) p_{\theta}(a \mid x)\right)\end{aligned}\]


by moving to the joint space…

  • allows us to learn complex approximating distn \(q_{\phi}(x,a)\) ,

    whose marginal \(q_{\phi}(x)=\int q_{\phi}(x \mid a) q_{\phi}(a) d a\) may be intractable


2-3. Combining Auxiliary VI & MCMC

after fitting variational approximation to \(p(x)\)… have 3 variational distns

  • 1) \(q_{*}(x \mid a)\).
  • 2) \(q_{*}(a)\).
  • 3) \(p_{*}(a \mid x)\).

These 1)~3) satisfies…

  • a) \(\int p(x) p_{*}(a \mid x) d x \approx q_{*}(a)\).

  • b) \(\int q_{*}(x \mid a) q_{*}(a) d a \approx p(x)\).

    ( becomes exact, if divergence \(\mathrm{KL}\left(q_{\phi}(x, a) \mid \mid p(x) p_{\theta}(a \mid x)\right)\) becomes zero )

(1) Naive Proposal Distributions

Example 1) Simple proposal

  • proposal ( \(\tilde{q}\) ) : constructed from variational distn \(q_{*}\)

  • perform MH-algorithm in the joint \((x, a)\) space,

    with proposal \(\tilde{q}\left(x^{\prime}, a^{\prime} \mid x, a\right)=q_{*}\left(x^{\prime} \mid a^{\prime}\right) q_{*}\left(a^{\prime}\right)\)

  • If variational approximation is accurate ( \(q(x, a) \approx p(x) p(a \mid x)\) ), HIGH acceptance rate


Example 2)

  • proposal : \(\tilde{q}\left(x^{\prime} \mid x\right)=\int q_{*}\left(x^{\prime} \mid a\right) p_{*}(a \mid x) d a\)

  • \(\begin{aligned} \int \tilde{q}\left(x^{\prime} \mid x\right) p(x) d x &=\int q_{*}\left(x^{\prime} \mid a\right) p_{*}(a \mid x) p(x) d x d a \approx \int q_{*}\left(x^{\prime} \mid a\right) q_{*}(x \mid a) q_{*}(a) d x d a \\ &=\int q_{*}\left(x^{\prime} \mid a\right) q_{*}(a) d a \approx \int p_{*}\left(a \mid x^{\prime}\right) p_{*}(a) d a=p\left(x^{\prime}\right) \end{aligned}\).


(2) Auxiliary Random Walk Proposal Distribution

Example 3) Auxiliary Random Walk Proposal Distribution

  • additional random perturbation in the auxiliary \(a\)-space

  • proposal : \(\tilde{q}\left(x^{\prime} \mid x\right)=\int q_{*}\left(x^{\prime} \mid a^{\prime}\right) \tilde{q}\left(a^{\prime} \mid a\right) p_{*}(a \mid x) d a d a^{\prime}\).

    where \(\tilde{q}\left(a^{\prime} \mid a\right)=\mathcal{N}\left(a^{\prime} \mid a, \sigma_{a}^{2} I\right)\)

  • Summary

    • 1) mapping from high-dim \(x\) to low-dim \(a\)
    • 2) perform random walk in \(a\)-space
    • 3) mapping back to high-dim space \(x\)


figure2


2-4. Choosing the Variational Family

Not finished yet! Have to decide the structure of variational distn, \(q(a, x)\) and \(p(a \mid x)\)

\(\rightarrow\) use DNN!

  • \(q(a) =\mathcal{N}(a \mid 0, I)\).

  • \(q_{\phi}(x \mid a) =\mathcal{N}\left(x \mid \mu_{\phi}(a), \Sigma_{\phi}(a)\right)\).

  • \(p_{\theta}(a \mid x) =\mathcal{N}\left(a \mid \mu_{\theta}(x), \Sigma_{\theta}(x)\right)\).

    where \(q_{\phi}(x \mid a)\) and \(p_{\theta}(a \mid x)\) are diagonal Gaussian


Key to flexibility of the auxiliary variational method ?

\(\rightarrow\) while \(q(a,x)\) can be evaluated point-wise, the marginal \(q(x)\) is a much richer approximating distn


Algorithm Summary

figure2

Categories:

Updated: