Fork me on GitHub

2.1.3.2.1. Variational Gaussian Mixture Models

The API is identical to that of the GMM class, the main difference being that it offers access to precision matrices as well as covariance matrices.

The inference algorithm is the one from the following paper:

While this paper presents the parts of the inference algorithm that are concerned with the structure of the dirichlet process, it does not go into detail in the mixture modeling part, which can be just as complex, or even more. For this reason we present here a full derivation of the inference algorithm and all the update and lower-bound equations. If you’re not interested in learning how to derive similar algorithms yourself and you’re not interested in changing/debugging the implementation in the scikit this document is not for you.

The complexity of this implementation is linear in the number of mixture components and data points. With regards to the dimensionality, it is linear when using spherical or diag and quadratic/cubic when using tied or full. For spherical or diag it is O(n_states * n_points * dimension) and for tied or full it is O(n_states * n_points * dimension^2 + n_states * dimension^3) (it is necessary to invert the covariance/precision matrices and compute its determinant, hence the cubic term).

This implementation is expected to scale at least as well as EM for the mixture of Gaussians.

2.1.3.2.2. Update rules for VB inference

Here the full mathematical derivation of the Variational Bayes update rules for Gaussian Mixture Models is given. The main parameters of the model, defined for any class k \in [1..K] are the class proportion \phi_k, the mean parameters \mu_k, the covariance parameters \Sigma_k, which is characterized by variational Wishart density, Wishart(a_k, \mathbf{B_k}), where a is the degrees of freedom, and B is the scale matrix. Depending on the covariance parametrization, B_k can be a positive scalar, a positive vector or a Symmetric Positive Definite matrix.

2.1.3.2.2.1. The spherical model

The model then is

\begin{array}{rcl}
\phi_k   &\sim& Beta(1, \alpha_1) \\
\mu_k   &\sim& Normal(0,  \mathbf{I}) \\
\sigma_k &\sim& Gamma(1, 1) \\
z_{i}     &\sim& SBP(\phi) \\
X_t &\sim& Normal(\mu_{z_i}, \frac{1}{\sigma_{z_i}} \mathbf{I})
\end{array}

The variational distribution we’ll use is

\begin{array}{rcl}
\phi_k   &\sim& Beta(\gamma_{k,1}, \gamma_{k,2}) \\
\mu_k   &\sim& Normal(\nu_{\mu_k},  \mathbf{I}) \\
\sigma_k &\sim& Gamma(a_{k}, b_{k}) \\
z_{i}     &\sim& Discrete(\nu_{z_i}) \\
\end{array}

2.1.3.2.2.1.1. The bound

The variational bound is

\begin{array}{rcl}
\log P(X) &\ge&
\sum_k (E_q[\log P(\phi_k)] - E_q[\log Q(\phi_k)]) \\
&&
+\sum_k \left( E_q[\log P(\mu_k)] - E_q[\log Q(\mu_k)] \right) \\
&&
+\sum_k \left( E_q[\log P(\sigma_k)] - E_q[\log Q(\sigma_k)] \right) \\
&&
+\sum_i \left( E_q[\log P(z_i)] - E_q[\log Q(z_i)] \right) \\
&&
+\sum_i E_q[\log P(X_t)]
\end{array}

The bound for \phi_k

\begin{array}{rcl}
E_q[\log Beta(1,\alpha)] - E[\log Beta(\gamma_{k,1},\gamma_{k,2})]
&=&
\log \Gamma(1+\alpha) - \log \Gamma(\alpha) \\ &&
+(\alpha-1)(\Psi(\gamma_{k,2})-\Psi(\gamma_{k,1}+\gamma_{k,2})) \\ &&
- \log \Gamma(\gamma_{k,1}+\gamma_{k,2}) + \log \Gamma(\gamma_{k,1}) +
\log \Gamma(\gamma_{k,2}) \\ &&
-
(\gamma_{k,1}-1)(\Psi(\gamma_{k,1})-\Psi(\gamma_{k,1}+\gamma_{k,2}))
\\ &&
-
(\gamma_{k,2}-1)(\Psi(\gamma_{k,2})-\Psi(\gamma_{k,1}+\gamma_{k,2}))
\end{array}

The bound for \mu_k

\begin{array}{rcl}
&& E_q[\log P(\mu_k)] - E_q[\log Q(\mu_k)] \\
&=&
\int\!d\mu_f q(\mu_f) \log P(\mu_f)
- \int\!d\mu_f q(\mu_f) \log Q(\mu_f)  \\
&=&
- \frac{D}{2}\log 2\pi - \frac{1}{2} ||\nu_{\mu_k}||^2 - \frac{D}{2}
+ \frac{D}{2} \log 2\pi e
\end{array}

The bound for \sigma_k

Here I’ll use the inverse scale parametrization of the gamma distribution.

\begin{array}{rcl}
&& E_q[\log P(\sigma_k)] - E_q [\log Q(\sigma_k)] \\ &=&
\log \Gamma (a_k) - (a_k-1)\Psi(a_k) -\log b_k + a_k - \frac{a_k}{b_k}
\end{array}

The bound for z

\begin{array}{rcl}
&& E_q[\log P(z)] - E_q[\log Q(z)] \\
&=&
\sum_{k} \left(
     \left(\sum_{j=k+1}^K  \nu_{z_{i,j}}\right)(\Psi(\gamma_{k,2})-\Psi(\gamma_{k,1}+\gamma_{k,2}))
 +  \nu_{z_{i,k}}(\Psi(\gamma_{k,1})-\Psi(\gamma_{k,1}+\gamma_{k,2}))
 - \log \nu_{z_{i,k}} \right)
\end{array}

The bound for X

Recall that there is no need for a Q(X) so this bound is just

\begin{array}{rcl}
E_q[\log P(X_i)] &=& \sum_k \nu_{z_k} \left( - \frac{D}{2}\log 2\pi
+\frac{D}{2} (\Psi(a_k) - \log(b_k))
-\frac{a_k}{2b_k} (||X_i - \nu_{\mu_k}||^2+D) - \log 2 \pi e  \right)
\end{array}

For simplicity I’ll later call the term inside the parenthesis E_q[\log P(X_i|z_i=k)]

2.1.3.2.2.1.2. The updates

Updating \gamma

\begin{array}{rcl}
\gamma_{k,1} &=& 1+\sum_i \nu_{z_{i,k}} \\
\gamma_{k,2} &=& \alpha + \sum_i \sum_{j > k} \nu_{z_{i,j}}.
\end{array}

Updating \mu

The updates for mu essentially are just weighted expectations of X regularized by the prior. We can see this by taking the gradient of the bound with regards to \nu_{\mu} and setting it to zero. The gradient is

\nabla L = -\nu_{\mu_k} + \sum_i \frac{\nu_{z_{i,k}}b_k}{a_k}(X_i + -\nu_{\mu})

so the update is

\nu_{\mu_k} = \frac{\sum_i \frac{\nu_{z_{i,k}}b_k}{a_k}X_i}{1+\sum_i \frac{\nu_{z_{i,k}}b_k}{a_k}}

Updating a and b

For some odd reason it doesn’t really work when you derive the updates for a and b using the gradients of the lower bound (terms involving the \Psi' function show up and a is hard to isolate). However, we can use the other formula,

\log Q(\sigma_k) = E_{v \ne \sigma_k}[\log P] + const

All the terms not involving \sigma_k get folded over into the constant and we get two terms: the prior and the probability of X. This gives us

\log Q(\sigma_k) = -\sigma_k  + \frac{D}{2} \sum_i \nu_{z_{i,k}}\log \sigma_k  - \frac{\sigma_k}{2}\sum_i \nu_{z_{i,k}} (||X_i-\mu_k||^2 + D)

This is the log of a gamma distribution, with a_k = 1+\frac{D}{2}\sum_i \nu_{z_{i,k}} and

b_k = 1 + \frac{1}{2}\sum_i \nu_{z_{i,k}} (||X_i-\mu_k||^2 + D).

You can verify this by normalizing the previous term.

Updating z

\log \nu_{z_{i,k}} \propto \Psi(\gamma_{k,1}) -
\Psi(\gamma_{k,1} + \gamma_{k,2}) + E_Q[\log P(X_i|z_i=k)] +
\sum_{j < k} \left (\Psi(\gamma_{j,2}) -
\Psi(\gamma_{j,1}+\gamma_{j,2})\right).

2.1.3.2.2.2. The diagonal model

The model then is

\begin{array}{rcl}
\phi_k   &\sim& Beta(1, \alpha_1) \\
\mu_k   &\sim& Normal(0,  \mathbf{I}) \\
\sigma_{k,d} &\sim& Gamma(1, 1) \\
z_{i}     &\sim& SBP(\phi) \\
X_t &\sim& Normal(\mu_{z_i}, \bm{\sigma_{z_i}}^{-1})
\end{array}

Tha variational distribution we’ll use is

\begin{array}{rcl}
\phi_k   &\sim& Beta(\gamma_{k,1}, \gamma_{k,2}) \\
\mu_k   &\sim& Normal(\nu_{\mu_k},  \mathbf{I}) \\
\sigma_{k,d} &\sim& Gamma(a_{k,d}, b_{k,d}) \\
z_{i}     &\sim& Discrete(\nu_{z_i}) \\
\end{array}

2.1.3.2.2.2.1. The lower bound

The changes in this lower bound from the previous model are in the distributions of \sigma (as there are a lot more \sigma s now) and X.

The bound for \sigma_{k,d} is the same bound for \sigma_k and can be safely omitted.

The bound for X :

The main difference here is that the precision matrix \bm{\sigma_k} scales the norm, so we have an extra term after computing the expectation of \mu_k^T\bm{\sigma_k}\mu_k, which is \nu_{\mu_k}^T\bm{\sigma_k}\nu_{\mu_k} + \sum_d \sigma_{k,d}. We then have

\begin{array}{rcl}
E_q[\log P(X_i)] &=& \sum_k \nu_{z_k} \Big( - \frac{D}{2}\log 2\pi
+\frac{1}{2}\sum_d (\Psi(a_{k,d}) - \log(b_{k,d})) \\
&&
-\frac{1}{2}((X_i - \nu_{\mu_k})^T\bm{\frac{a_k}{b_k}}(X_i - \nu_{\mu_k})+ \sum_d \sigma_{k,d})- \log 2 \pi e  \Big)
\end{array}

2.1.3.2.2.2.2. The updates

The updates only chance for \mu (to weight them with the new \sigma), z (but the change is all folded into the E_q[P(X_i|z_i=k)] term), and the a and b variables themselves.

The update for \mu

\nu_{\mu_k} = \left(\mathbf{I}+\sum_i \frac{\nu_{z_{i,k}}\mathbf{b_k}}{\mathbf{a_k}}\right)^{-1}\left(\sum_i \frac{\nu_{z_{i,k}}b_k}{a_k}X_i\right)

The updates for a and b

Here we’ll do something very similar to the spheric model. The main difference is that now each \sigma_{k,d} controls only one dimension of the bound:

\log Q(\sigma_{k,d}) = -\sigma_{k,d} + \sum_i \nu_{z_{i,k}}\frac{1}{2}\log \sigma_{k,d}
- \frac{\sigma_{k,d}}{2}\sum_i \nu_{z_{i,k}} ((X_{i,d}-\mu_{k,d})^2 + 1)

Hence

a_{k,d} = 1 + \frac{1}{2} \sum_i \nu_{z_{i,k}}

b_{k,d} = 1 + \frac{1}{2} \sum_i \nu_{z_{i,k}}((X_{i,d}-\mu_{k,d})^2 + 1)

2.1.3.2.2.3. The tied model

The model then is

\begin{array}{rcl}
\phi_k   &\sim& Beta(1, \alpha_1) \\
\mu_k   &\sim& Normal(0,  \mathbf{I}) \\
\Sigma &\sim& Wishart(D, \mathbf{I}) \\
z_{i}     &\sim& SBP(\phi) \\
X_t &\sim& Normal(\mu_{z_i},  \Sigma^{-1})
\end{array}

Tha variational distribution we’ll use is

\begin{array}{rcl}
\phi_k   &\sim& Beta(\gamma_{k,1}, \gamma_{k,2}) \\
\mu_k   &\sim& Normal(\nu_{\mu_k},  \mathbf{I}) \\
\Sigma &\sim& Wishart(a, \mathbf{B}) \\
z_{i}     &\sim& Discrete(\nu_{z_i}) \\
\end{array}

2.1.3.2.2.3.1. The lower bound

There are two changes in the lower-bound: for \Sigma and for X.

The bound for \Sigma

\begin{array}{rcl}
\frac{D^2}{2}\log 2  + \sum_d \log \Gamma(\frac{D+1-d}{2}) \\
- \frac{aD}{2}\log 2 + \frac{a}{2} \log |\mathbf{B}| + \sum_d \log \Gamma(\frac{a+1-d}{2}) \\
+ \frac{a-D}{2}\left(\sum_d \Psi\left(\frac{a+1-d}{2}\right)
+ D \log 2 + \log |\mathbf{B}|\right) \\
+ \frac{1}{2} a \mathbf{tr}[\mathbf{B}-\mathbf{I}]
\end{array}

The bound for X

\begin{array}{rcl}
E_q[\log P(X_i)] &=& \sum_k \nu_{z_k} \Big( - \frac{D}{2}\log 2\pi
+\frac{1}{2}\left(\sum_d \Psi\left(\frac{a+1-d}{2}\right)
+ D \log 2 + \log |\mathbf{B}|\right) \\
&&
-\frac{1}{2}((X_i - \nu_{\mu_k})a\mathbf{B}(X_i - \nu_{\mu_k})+ a\mathbf{tr}(\mathbf{B}))- \log 2 \pi e  \Big)
\end{array}

2.1.3.2.2.3.2. The updates

As in the last setting, what changes are the trivial update for z, the update for \mu and the update for a and \mathbf{B}.

The update for \mu

\nu_{\mu_k} = \left(\mathbf{I}+ a\mathbf{B}\sum_i \nu_{z_{i,k}}\right)^{-1}
\left(a\mathbf{B}\sum_i \nu_{z_{i,k}} X_i\right)

The update for a and B

As this distribution is far too complicated I’m not even going to try going at it the gradient way.

\log Q(\Sigma) = +\frac{1}{2}\log |\Sigma| - \frac{1}{2} \mathbf{tr}[\Sigma]
+ \sum_i \sum_k \nu_{z_{i,k}} \left( +\frac{1}{2}\log |\Sigma| - \frac{1}{2}((X_i-\nu_{\mu_k})^T\Sigma(X_i-\nu_{\mu_k})+\mathbf{tr}[\Sigma]) \right)

which non-trivially (seeing that the quadratic form with \Sigma in the middle can be expressed as the trace of something) reduces to

\log Q(\Sigma) = +\frac{1}{2}\log |\Sigma| - \frac{1}{2} \mathbf{tr}[\Sigma]
+ \sum_i \sum_k \nu_{z_{i,k}} \left( +\frac{1}{2}\log |\Sigma| - \frac{1}{2}(\mathbf{tr}[(X_i-\nu_{\mu_k})(X_i-\nu_{\mu_k})^T\Sigma]+\mathbf{tr}[I \Sigma]) \right)

hence this (with a bit of squinting) looks like a wishart with parameters

a = 2 + D + T

and

\mathbf{B} = \left(\mathbf{I} + \sum_i \sum_k \nu_{z_{i,k}}(X_i-\nu_{\mu_k})(X_i-\nu_{\mu_k})^T\right)^{-1}

2.1.3.2.2.4. The full model

The model then is

\begin{array}{rcl}
\phi_k   &\sim& Beta(1, \alpha_1) \\
\mu_k   &\sim& Normal(0,  \mathbf{I}) \\
\Sigma_k &\sim& Wishart(D, \mathbf{I}) \\
z_{i}     &\sim& SBP(\phi) \\
X_t &\sim& Normal(\mu_{z_i},  \Sigma_{z,i}^{-1})
\end{array}

The variational distribution we’ll use is

\begin{array}{rcl}
\phi_k   &\sim& Beta(\gamma_{k,1}, \gamma_{k,2}) \\
\mu_k   &\sim& Normal(\nu_{\mu_k},  \mathbf{I}) \\
\Sigma_k &\sim& Wishart(a_k, \mathbf{B_k}) \\
z_{i}     &\sim& Discrete(\nu_{z_i}) \\
\end{array}

2.1.3.2.2.4.1. The lower bound

All that changes in this lower bound in comparison to the previous one is that there are K priors on different \Sigma precision matrices and there are the correct indices on the bound for X.

2.1.3.2.2.4.2. The updates

All that changes in the updates is that the update for mu uses only the proper sigma and the updates for a and B don’t have a sum over K, so

\nu_{\mu_k} = \left(\mathbf{I}+ a_k\mathbf{B_k}\sum_i \nu_{z_{i,k}}\right)^{-1}
\left(a_k\mathbf{B_k}\sum_i \nu_{z_{i,k}} X_i\right)

a_k = 2 + D + \sum_i \nu_{z_{i,k}}

and

\mathbf{B} = \left(\left(\sum_i\nu_{z_{i,k}}+1\right)\mathbf{I} + \sum_i  \nu_{z_{i,k}}(X_i-\nu_{\mu_k})(X_i-\nu_{\mu_k})^T\right)^{-1}

Previous
Next