User:Timothee Flutre/Notebook/Postdoc/2012/08/16
From OpenWetWare
(→Entry title: first version) |
(→Variational Bayes approach for the mixture of Normals: finish update q_w) |
||
(10 intermediate revisions not shown.) | |||
Line 11: | Line 11: | ||
- | * '''Data''': | + | * '''Data''': N univariate observations, <math>y_1, \ldots, y_N</math>, gathered into the vector <math>\mathbf{y}</math> |
+ | * '''Model''': mixture of K Normal distributions | ||
- | * ''' | + | * '''Parameters''': K mixture weights (<math>w_k</math>), K means (<math>\mu_k</math>) and K [http://en.wikipedia.org/wiki/Precision_%28statistics%29 precisions] (<math>\tau_k</math>), one per mixture component |
+ | <math>\Theta = \{w_1,\ldots,w_K,\mu_1,\ldots,\mu_K,\tau_1,\ldots,\tau_K\}</math> | ||
- | * ''' | + | * '''Constraints''': <math>\sum_{k=1}^K w_k = 1</math> and <math>\forall k \; w_k > 0</math>. |
+ | * '''Observed likelihood''': observations assumed exchangeable (independent and identically distributed given the parameters) | ||
- | + | <math>p(\mathbf{y} | \Theta, K) = \prod_{n=1}^N p(y_n|\Theta,K) = \prod_{n=1}^N \sum_{k=1}^K w_k \; \mathcal{N}(y_n;\mu_k,\tau_k^{-1})</math> | |
+ | * '''Latent variables''': N hidden variables, <math>z_1,\ldots,z_N</math>, each being a vector of length K with a single 1 indicating the component to which the <math>n^{th}</math> observation belongs, and K-1 zeroes. | ||
- | + | <math>p(\mathbf{z}|\mathbf{w},K) = \prod_{n=1}^N p(z_n|\mathbf{w},K) = \prod_{n=1}^N \prod_{k=1}^K w_k^{z_{nk}}</math> | |
+ | * '''Augmented likelihood''': | ||
- | + | <math>p(\mathbf{y},\mathbf{z}|\Theta,K) = \prod_{n=1}^N p(y_n,z_n|\Theta,K) = \prod_{n=1}^N p(z_n|\Theta,K) p(y_n|z_n,\Theta,K) = \prod_{n=1}^N \prod_{k=1}^K w_k^{z_{nk}} \; \mathcal{N}(y_n;\mu_k,\tau_k^{-1})^{z_{nk}}</math> | |
- | + | ||
- | + | ||
+ | * '''Maximum-likelihood estimation''': integrate out the latent variables | ||
+ | |||
+ | <math>\mathrm{ln} \, p(\mathbf{y} | \Theta, K) = \sum_{n=1}^N \mathrm{ln} \, \int_{z_n} p(y_n, z_n | \Theta, K) \, \mathrm{d}{z_n}</math> | ||
+ | |||
+ | The latent variables induce dependencies between all the parameters of the model. | ||
+ | This makes it difficult to find the parameters that maximize the likelihood. | ||
+ | An elegant solution is to introduce a variational distribution of parameters and latent variables, which leads to a re-formulation of the classical EM algorithm. | ||
+ | But let's show it directly in the Bayesian paradigm. | ||
+ | |||
+ | * '''Priors''': conjuguate | ||
+ | ** <math>\forall k \; \tau_k \sim \mathcal{G}a(\alpha,\beta)</math> and <math>\forall k \; \mu_k | \tau_k \sim \mathcal{N}(\mu_0,(\tau_0 \tau_k)^{-1})</math> | ||
+ | ** <math>\forall n \; z_n \sim \mathcal{M}ult_K(1,\mathbf{w})</math> and <math>\mathbf{w} \sim \mathcal{D}ir(\gamma_0)</math> | ||
+ | |||
+ | |||
+ | * '''Variational Bayes''': let's focus here on calculating the marginal log-likelihood of our data set in order to perform model comparison: | ||
+ | |||
+ | <math>\mathrm{ln} \, p(\mathbf{y} | K) = \mathrm{ln} \, \int_\mathbf{z} \int_\Theta \; p(\mathbf{y}, \mathbf{z}, \Theta | K) \, \mathrm{d}\mathbf{z} \, \mathrm{d}\Theta</math> | ||
+ | |||
+ | We can now introduce a distribution <math>q_{\mathbf{z}, \Theta}</math>: | ||
+ | |||
+ | <math>\mathrm{ln} \, p(\mathbf{y} | K) = \mathrm{ln} \, \left( \int_\mathbf{z} \int_\Theta \; q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta) \; \frac{p(\mathbf{y}, \mathbf{z}, \Theta | K)}{q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta)} \, \mathrm{d}\mathbf{z} \, \mathrm{d}\Theta \right) + C_{\mathbf{z}, \Theta}</math> | ||
+ | |||
+ | The constant <math>C_{\mathbf{z}, \Theta}</math> is here to remind us that <math>q_{\mathbf{z}, \Theta}</math> has the constraint of being a distribution, ie. of summing to 1, which can be enforced by a Lagrange multiplier. | ||
+ | |||
+ | We can then use the concavity of the logarithm ([http://en.wikipedia.org/wiki/Jensen%27s_inequality Jensen's inequality]) to derive a lower bound of the marginal log-likelihood: | ||
+ | |||
+ | <math>\mathrm{ln} \, p(\mathbf{y} | K) \ge \int_\mathbf{z} \int_\Theta \; q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta) \; \mathrm{ln} \, \frac{p(\mathbf{y}, \mathbf{z}, \Theta | K)}{q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta)} \, \mathrm{d}\mathbf{z} \, \mathrm{d}\Theta + C_{\mathbf{z}, \Theta} = \mathcal{F}_K(q)</math> | ||
+ | |||
+ | Let's call this lower bound <math>\mathcal{F}_K(q)</math> as it is a [http://en.wikipedia.org/wiki/Functional_%28mathematics%29 functional], ie. a ''function of functions''. To gain some intuition about the impact of introducing <math>q</math>, let's expand <math>\mathcal{F}_K</math>: | ||
+ | |||
+ | <math>\mathcal{F}_K(q) = \int_\mathbf{z} \int_\Theta \; q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta) \; \mathrm{ln} \, p(\mathbf{y} | \mathbf{z}, \Theta, K) \, \mathrm{d}\mathbf{z} \, \mathrm{d}\Theta \; + \; \int_\mathbf{z} \int_\Theta \; q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta) \; \mathrm{ln} \, \frac{p(\mathbf{z}, \Theta | K)}{q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta)} \, \mathrm{d}\mathbf{z} \, \mathrm{d}\Theta \; + \; C_{\mathbf{z}, \Theta}</math> | ||
+ | |||
+ | <math>\mathcal{F}_K(q) = \mathrm{ln} \, p(\mathbf{y} | \mathbf{z}, \Theta, K) - D_{KL}(q || p)</math> | ||
+ | |||
+ | From this, it is clear that <math>\mathcal{F}_K</math> (ie. a lower-bound of the marginal log-likelihood) is the conditional log-likelihood minus the [http://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence Kullback-Leibler divergence] between the variational distribution <math>q</math> and the joint posterior of latent variables and parameters. | ||
+ | As a side note, minimizing <math>D_{KL}(p || q)</math> is used in the [http://en.wikipedia.org/wiki/Expectation_propagation expectation-propagation] technique. | ||
+ | |||
+ | In practice, we have to make the following crucial assumption of independence on <math>q_{\mathbf{z}, \Theta}</math> in order for the calculations to be analytically tractable: | ||
+ | |||
+ | <math>q_{\mathbf{z}, \Theta}(\mathbf{z}, \Theta) = q_{\mathbf{z}}(\mathbf{z}) q_\Theta(\Theta)</math> | ||
+ | |||
+ | This means that <math>q_\mathbf{z} q_\Theta</math> approximates the joint posterior, and therefore the lower-bound will be tight if and only if this approximation is exact and the KL divergence is zero. | ||
+ | |||
+ | As we ultimately aim at inferring the parameters and latent variables that maximize the marginal log-likelihood, we will use the [http://en.wikipedia.org/wiki/Calculus_of_variations calculus of variations] to find the functions <math>q_\mathbf{z}</math> and <math>q_\Theta</math> that maximize the functional <math>\mathcal{F}_K</math>. | ||
+ | |||
+ | <math>\mathcal{F}_K(q_\mathbf{z}, q_\Theta) = \int_\Theta \; q_\Theta(\Theta) \; \left( \int_\mathbf{z} \; q_\mathbf{z}(\mathbf{z}) \; \mathrm{ln} \, \frac{p(\mathbf{y}, \mathbf{z} | \Theta, K)}{q_\mathbf{z}(\mathbf{z})} \, \mathrm{d}\mathbf{z} + \mathrm{ln} \, \frac{p(\Theta | K)}{q_\Theta(\Theta)} \right) \, \mathrm{d}\Theta \; + \; C_{\mathbf{z}} \; + \; C_{\Theta}</math> | ||
+ | |||
+ | This naturally leads to a procedure very similar to the EM algorithm where, at the E step, we calculate the expectations of the parameters with respect to the variational distributions <math>q_\mathbf{z}</math> and <math>q_\Theta</math>, and, at the M step, we recompute the variational distributions over the parameters. | ||
+ | |||
+ | |||
+ | * '''Updates for <math>q_\mathbf{z}</math>''': | ||
+ | |||
+ | We start by writing the functional derivative of <math>\mathcal{F}_K</math> with respect to <math>q_{\mathbf{z}}</math>: | ||
+ | |||
+ | <math>\frac{\partial \mathcal{F}_K}{\partial q_{\mathbf{z}}} = \int_\Theta \; q_\Theta(\Theta) \; \frac{\partial}{\partial q_{\mathbf{z}}} \left( \int_\mathbf{z} \; \left( q_{\mathbf{z}}(\mathbf{z}) \mathrm{ln} \, p(\mathbf{y},\mathbf{z}|\Theta,K) - q_{\mathbf{z}}(\mathbf{z}) \mathrm{ln} \, q_{\mathbf{z}}(\mathbf{z}) \right) \, \mathrm{d}\mathbf{z} \right) \, \mathrm{d}\Theta \; + \; C_{\mathbf{z}}</math> | ||
+ | |||
+ | <math>\frac{\partial \mathcal{F}_K}{\partial q_{\mathbf{z}}} = \int_\Theta \; q_\Theta(\Theta) \; \left( \mathrm{ln} \, p(\mathbf{y},\mathbf{z}|\Theta,K) - \mathrm{ln} \, q_{\mathbf{z}}(\mathbf{z}) - 1 \right) \, \mathrm{d}\Theta \; + \; C_{\mathbf{z}}</math> | ||
+ | |||
+ | Then we set this functional derivative to zero. We also make use of a frequent assumption, namely that the variational distribution fully factorizes over each individual latent variables ([http://en.wikipedia.org/wiki/Mean_field_theory mean-field assumption]): | ||
+ | |||
+ | <math>\frac{\partial \mathcal{F}_K}{\partial q_{\mathbf{z}}} \bigg|_{q_{\mathbf{z}}^{(t+1)}} = 0 \Longleftrightarrow \forall \, n \; \mathrm{ln} \, q_{z_n}^{(t+1)}(z_n) = \int_\Theta \; q_\Theta(\Theta) \; \mathrm{ln} \, p(y_n,z_n|\Theta,K) \, \mathrm{d}\Theta \; - \; 1 \; + \; C_{z_n}</math> | ||
+ | |||
+ | Recognizing the expectation and factorizing <math>q_\Theta(\Theta)</math> into <math>q_\mathbf{w}(\mathbf{w})q_\mathbf{\mu,\tau}(\mathbf{\mu,\tau})</math>, we get: | ||
+ | |||
+ | <math>\mathrm{ln} \, q_{z_n}^{(t+1)}(z_n) = \mathbb{E}_\mathbf{w}[\mathrm{ln} \, p(z_n|\mathbf{w},K)] + \mathbb{E}_{\mathbf{\mu,\tau}}[\mathrm{ln} \, p(y_n|z_n,\mathbf{\mu},\mathbf{\tau},K)] \; + \; \text{constant}</math> | ||
+ | |||
+ | <math>\mathrm{ln} \, q_{z_n}^{(t+1)}(z_n) = \sum_{k=1}^K ( z_{nk} \; \mathrm{ln} \, \rho_{nk} ) \; + \; \text{constant}</math> where <math>\mathrm{ln} \, \rho_{nk} = \mathbb{E}[\mathrm{ln} \, w_k] + \frac{1}{2} \mathbb{E}[\mathrm{ln} \, \tau_k] - \frac{1}{2} \mathrm{ln} \, 2\pi - \frac{1}{2} \mathbb{E}[\tau_k (y_n-\mu_k)^2]</math> | ||
+ | |||
+ | Taking the exponential: <math>q_{z_n}^{(t+1)}(z_n) \propto \prod_k \rho^{z_{nk}}</math> | ||
+ | |||
+ | As this should be a distribution, it should sum to one, and therefore: | ||
+ | |||
+ | <math>q_{z_n}^{(t+1)}(z_n) = \prod_k r_{nk}^{z_{nk}}</math> where <math>r_{nk} = \frac{\rho_{nk}}{\sum_{k'=1}^K \rho_{nk'}}</math> ("r" stands for "reponsability") | ||
+ | |||
+ | Interestingly, even though we haven't specified anything yet about <math>q_{z_n}</math>, we can see that it is of the same form as the prior on <math>z_n</math>, a Multinomial distribution. | ||
+ | |||
+ | |||
+ | * '''Updates for <math>q_\Theta</math>''': | ||
+ | |||
+ | We start by writing the functional derivative of <math>\mathcal{F}_K</math> with respect to <math>q_{\Theta}</math>: | ||
+ | |||
+ | <math>\frac{\partial \mathcal{F}_K}{\partial q_\Theta} = \frac{\partial}{\partial q_\Theta} \left( \int_\Theta \; q_\Theta(\Theta) \; \left( \int_\mathbf{z} \; q_\mathbf{z}(\mathbf{z}) \; \mathrm{ln} \, p(\mathbf{y}, \mathbf{z} | \Theta, K) \, \mathrm{d}\mathbf{z} + \mathrm{ln} \, \frac{p(\Theta | K)}{q_\Theta(\Theta)} \right) \, \mathrm{d}\Theta \right) \; + \; C_{\Theta}</math> | ||
+ | |||
+ | <math>\frac{\partial \mathcal{F}_K}{\partial q_\Theta} = \int_\mathbf{z} \; q_\mathbf{z}(\mathbf{z}) \; \mathrm{ln} \, p(\mathbf{y}, \mathbf{z} | \Theta, K) \, \mathrm{d}\mathbf{z} + \mathrm{ln} \, p(\Theta | K) - \mathrm{ln} \, q_\Theta(\Theta) \; + \; C_{\Theta}</math> | ||
+ | |||
+ | Then, when setting this functional derivative to zero and using the factorization <math>q_\Theta = q_\mathbf{w}(\mathbf{w})q_\mathbf{\mu,\tau}(\mathbf{\mu,\tau})</math>, we can obtain the variational distribution of each parameter. | ||
+ | |||
+ | Starting with <math>\mathbf{w}</math>: | ||
+ | |||
+ | <math>\mathrm{ln} \, q_\mathbf{w}^{(t+1)}(\mathbf{w}) = \mathrm{ln} \, p(\mathbf{w}|K) + \mathbb{E}_\mathbf{z}[\mathrm{ln} \, p(\mathbf{z}|\mathbf{w},K)] \; + \; \text{constant}</math> | ||
+ | |||
+ | <math>\mathrm{ln} \, q_\mathbf{w}^{(t+1)}(\mathbf{w}) = (\gamma_0 - 1) \sum_k \mathrm{ln} \, w_k + \sum_n \sum_k \mathbb{E}[z_{nk}] \, \mathrm{ln} \, w_k \; + \; \text{constant}</math> | ||
+ | |||
+ | Recognizing <math>\mathbb{E}[z_{nk}] = r_{nk}</math> and taking the exponential, we get another Dirichlet distribution: | ||
+ | |||
+ | <math>q_\mathbf{w}^{(t+1)}(\mathbf{w}) \propto \prod_k w_k^{\gamma_0-1+\sum_n r_{nk}}</math> | ||
+ | |||
+ | that is <math>q_\mathbf{w}^{(t+1)} \sim \mathcal{D}ir(\gamma^{(t+1)})</math> with <math>\gamma_k^{(t+1)}=\gamma_0-1+\sum_n r_{nk}</math> | ||
+ | |||
+ | TODO | ||
+ | |||
+ | |||
+ | * '''Choice of K''': | ||
+ | |||
+ | TODO | ||
<!-- ##### DO NOT edit below this line unless you know what you are doing. ##### --> | <!-- ##### DO NOT edit below this line unless you know what you are doing. ##### --> |
Revision as of 20:53, 26 May 2013
Project name | Main project page Previous entry Next entry |
Variational Bayes approach for the mixture of Normals
The latent variables induce dependencies between all the parameters of the model. This makes it difficult to find the parameters that maximize the likelihood. An elegant solution is to introduce a variational distribution of parameters and latent variables, which leads to a re-formulation of the classical EM algorithm. But let's show it directly in the Bayesian paradigm.
We can now introduce a distribution :
The constant is here to remind us that has the constraint of being a distribution, ie. of summing to 1, which can be enforced by a Lagrange multiplier. We can then use the concavity of the logarithm (Jensen's inequality) to derive a lower bound of the marginal log-likelihood:
Let's call this lower bound as it is a functional, ie. a function of functions. To gain some intuition about the impact of introducing q, let's expand :
From this, it is clear that (ie. a lower-bound of the marginal log-likelihood) is the conditional log-likelihood minus the Kullback-Leibler divergence between the variational distribution q and the joint posterior of latent variables and parameters. As a side note, minimizing D_{KL}(p | | q) is used in the expectation-propagation technique. In practice, we have to make the following crucial assumption of independence on in order for the calculations to be analytically tractable:
This means that approximates the joint posterior, and therefore the lower-bound will be tight if and only if this approximation is exact and the KL divergence is zero. As we ultimately aim at inferring the parameters and latent variables that maximize the marginal log-likelihood, we will use the calculus of variations to find the functions and q_{Θ} that maximize the functional .
This naturally leads to a procedure very similar to the EM algorithm where, at the E step, we calculate the expectations of the parameters with respect to the variational distributions and q_{Θ}, and, at the M step, we recompute the variational distributions over the parameters.
We start by writing the functional derivative of with respect to :
Then we set this functional derivative to zero. We also make use of a frequent assumption, namely that the variational distribution fully factorizes over each individual latent variables (mean-field assumption):
Recognizing the expectation and factorizing q_{Θ}(Θ) into , we get:
where Taking the exponential: As this should be a distribution, it should sum to one, and therefore: where ("r" stands for "reponsability") Interestingly, even though we haven't specified anything yet about , we can see that it is of the same form as the prior on z_{n}, a Multinomial distribution.
We start by writing the functional derivative of with respect to q_{Θ}:
Then, when setting this functional derivative to zero and using the factorization , we can obtain the variational distribution of each parameter. Starting with :
Recognizing and taking the exponential, we get another Dirichlet distribution:
that is with TODO
TODO |