Inference Schemes ¶ Source Sampling
Advances in VI - Notebook
Numerical Integration (low dimension)
Bayesian Quadrature
Expectation Propagation
Conjugate Priors (Gaussian Likelihood w/ GP Prior)
Subset Methods (Nystrom)
Fast Linear Algebra (Krylov, Fast Transforms, KD-Trees)
Variational Methods (Laplace, Mean-Field, Expectation Propagation)
Monte Carlo Methods (Gibbs, Metropolis-Hashings, Particle Filter)
Local Methods
Sampling Methods
Local Methods ¶ Mean Squared Error (MSE) ¶ In the case of regression, we can use the MSE as a loss function. This will exactly solve for the negative log-likelihood term above.
The likelihood of our model is:
log p ( y ∣ X , w ) = ∑ i = 1 N log p ( y i ∣ x i , θ ) \log p(y|\mathbf{X,w}) = \sum_{i=1}^N \log p(y_i|x_i,\theta) log p ( y ∣ X , w ) = i = 1 ∑ N log p ( y i ∣ x i , θ ) And for simplicity, we assume the noise ϵ \epsilon ϵ
log p ( y ∣ X , w ) = ∑ i = 1 N log N ( y i ∣ x i w , σ 2 ) \log p(y|\mathbf{X,w}) = \sum_{i=1}^N \log \mathcal{N}(y_i | \mathbf{x}_i\mathbf{w}, \sigma^2) log p ( y ∣ X , w ) = i = 1 ∑ N log N ( y i ∣ x i w , σ 2 ) Plugging in the full formula for the Gaussian distribution with some simplifications gives us:
log p ( y ∣ X , w ) = ∑ i = 1 N log 1 2 π σ e 2 exp ( − ( y i − x i w ) 2 2 σ e 2 ) \log p(y|\mathbf{X,w}) =
\sum_{i=1}^N
\log \frac{1}{\sqrt{2 \pi \sigma_e^2}}
\exp\left( - \frac{(y_i - \mathbf{x}_i\mathbf{w})^2}{2\sigma_e^2} \right) log p ( y ∣ X , w ) = i = 1 ∑ N log 2 π σ e 2 1 exp ( − 2 σ e 2 ( y i − x i w ) 2 ) We can use the log rule log a b = log a + log b \log ab = \log a + \log b log ab = log a + log b log e x = x \log e^x = x log e x = x
log p ( y ∣ X , w ) = − N 2 log 2 π σ e 2 − ∑ i = 1 N ( y i − x i w ) 2 2 σ e 2 \log p(y|\mathbf{X,w}) = - \frac{N}{2} \log 2 \pi \sigma_e^2 - \sum_{i=1}^N \frac{(y_i - \mathbf{x_iw})^2}{2\sigma_e^2} log p ( y ∣ X , w ) = − 2 N log 2 π σ e 2 − i = 1 ∑ N 2 σ e 2 ( y i − x i w ) 2 So, the first term is constant so that we can ignore that in our loss function. We can do the same for the denominator for the second term. Let’s simplify it to make our life easier.
log p ( y ∣ X , w ) = − ∑ i = 1 N ( y i − x i w ) 2 \log p(y|\mathbf{X,w}) = - \sum_{i=1}^N (y_i - \mathbf{x}_i\mathbf{w})^2 log p ( y ∣ X , w ) = − i = 1 ∑ N ( y i − x i w ) 2 So we want to maximize this quantity: in other words, I want to find the parameter w \mathbf{w} w
w M L E = argmin w − ∑ i = 1 N ( y i − x i w ) 2 \mathbf{w}_{MLE} = \operatorname*{argmin}_{\mathbf{w}} - \sum_{i=1}^N (y_i - \mathbf{x}_i\mathbf{w})^2 w M L E = w argmin − i = 1 ∑ N ( y i − x i w ) 2 We can rewrite this expression because the maximum of a negative quantity is the same as minimizing a positive quantity.
w M L E = argmin w 1 N ∑ i = 1 N ( y i − x i w ) 2 \mathbf{w}_{MLE} = \operatorname*{argmin}_{\mathbf{w}} \frac{1}{N} \sum_{i=1}^N (y_i - \mathbf{x}_i\mathbf{w})^2 w M L E = w argmin N 1 i = 1 ∑ N ( y i − x i w ) 2 This is the same as the MSE error expression; with the edition of a scalar value 1 / N 1/N 1/ N
w M L E = argmin w 1 N ∑ i = 1 N ( y i − x i w ) 2 = argmin w MSE \begin{aligned}
\mathbf{w}_{MLE} &= \operatorname*{argmin}_{\mathbf{w}} \frac{1}{N} \sum_{i=1}^N (y_i - \mathbf{x}_i\mathbf{w})^2 \\
&= \operatorname*{argmin}_{\mathbf{w}} \text{MSE}
\end{aligned} w M L E = w argmin N 1 i = 1 ∑ N ( y i − x i w ) 2 = w argmin MSE Note : If we did not know σ y 2 \sigma_y^2 σ y 2
Sources :
Maximum A Priori (MAP) ¶ Loss Function ¶ θ MAP = argmax θ − 1 N ∑ n N log p ( y n ∣ f ( x n ; θ ) ) + log p ( θ ) \boldsymbol{\theta}_{\text{MAP}} = \operatorname*{argmax}_{\boldsymbol{\theta}} - \frac{1}{N}\sum_n^N\log p\left(y_n|f(x_n; \theta)\right) + \log p(\theta) θ MAP = θ argmax − N 1 n ∑ N log p ( y n ∣ f ( x n ; θ ) ) + log p ( θ )
θ MAP = argmax θ log p ( θ ∣ D ) \boldsymbol{\theta}_{\text{MAP}} = \operatorname*{argmax}_{\boldsymbol{\theta}} \log p(\boldsymbol{\theta}|\mathcal{D}) θ MAP = θ argmax log p ( θ ∣ D ) We can plug in the base Bayesian formulation
θ MAP = argmax θ log [ p ( D ∣ θ ) p ( θ ) p ( D ) ] \boldsymbol{\theta}_{\text{MAP}} = \operatorname*{argmax}_{\boldsymbol{\theta}} \log \left[ \frac{p(\mathcal{D}|\boldsymbol{\theta})p(\boldsymbol{\theta})}{p(\mathcal{D})} \right] θ MAP = θ argmax log [ p ( D ) p ( D ∣ θ ) p ( θ ) ] We can expand this term using the log rules
θ m a p = argmax θ [ log p ( D ∣ θ ) + log p ( θ ) − log p ( D ) ] \theta_{map} = \operatorname*{argmax}_\theta \left[ \log p(D|\theta) + \log p(\theta) - \log p(D) \right] θ ma p = θ argmax [ log p ( D ∣ θ ) + log p ( θ ) − log p ( D ) ] Notice that log p ( D ) \log p(D) log p ( D ) θ \theta θ
θ m a p = argmax θ [ log p ( D ∣ θ ) + log p ( θ ) ] \theta_{map} = \operatorname*{argmax}_\theta \left[ \log p(D|\theta) + \log p(\theta) \right] θ ma p = θ argmax [ log p ( D ∣ θ ) + log p ( θ ) ] We will change this problem into a minimization problem instead of maximization
θ m a p = argmin θ − log p ( D ∣ θ ) \theta_{map} = \operatorname*{argmin}_{\theta} - \log p(D|\theta) θ ma p = θ argmin − log p ( D ∣ θ ) We cannot find the probability distribution of p ( D ∣ θ ) p(D|\theta) p ( D ∣ θ )
θ m a p = argmin θ − E x ∼ P X [ log p ( D ∣ θ ) ] + log p ( θ ) \theta_{map} = \operatorname*{argmin}_{\theta} - \mathbb{E}_{\mathbf{x}\sim P_X} \left[ \log p(D|\theta)\right] + \log p(\theta) θ ma p = θ argmin − E x ∼ P X [ log p ( D ∣ θ ) ] + log p ( θ ) We can approximate this using Monte carlo samples. This is given by:
E x [ log p ( D ∣ θ ) ] ≈ 1 N ∑ n N p ( y n ∣ f ( x n ; θ ) ) \mathbb{E}_{x}[\log p(D|\theta)] \approx \frac{1}{N}\sum_n^N p(y_n | f(x_n; \theta)) E x [ log p ( D ∣ θ )] ≈ N 1 n ∑ N p ( y n ∣ f ( x n ; θ )) and we assume that with enough samples, we will capture the essence of our data.
θ m a p = argmin θ − 1 N ∑ n N log p ( y n ∣ f ( x n ; θ ) ) + log p ( θ ) \theta_{map} = \operatorname*{argmin}_{\theta} - \frac{1}{N}\sum_n^N \log p(y_n| f(x_n;\theta))+ \log p(\theta) θ ma p = θ argmin − N 1 n ∑ N log p ( y n ∣ f ( x n ; θ )) + log p ( θ ) Maximum Likelihood Estimation (MLE) ¶ Loss Function ¶ θ m a p = argmin θ − 1 N ∑ n N log p ( y n ∣ f ( x n ; θ ) ) \theta_{map} = \operatorname*{argmin}_{\theta} - \frac{1}{N}\sum_n^N \log p(y_n| f(x_n;\theta)) θ ma p = θ argmin − N 1 n ∑ N log p ( y n ∣ f ( x n ; θ ))
This is straightforward to derive because we can pick up from the proof of the MAP loss function, eq:(9)
θ m a p = argmin θ − 1 N ∑ n N log p ( y n ∣ f ( x n ; θ ) ) + log p ( θ ) \theta_{map} = \operatorname*{argmin}_{\theta} - \frac{1}{N}\sum_n^N \log p(y_n| f(x_n;\theta))+ \log p(\theta) θ ma p = θ argmin − N 1 n ∑ N log p ( y n ∣ f ( x n ; θ )) + log p ( θ ) In this case, we will assume a uniform prior on our parameters, θ \theta θ log \log log p ( θ ) = 1 p(\theta)=1 p ( θ ) = 1
θ m a p = argmin θ − 1 N ∑ n N log p ( y n ∣ f ( x n ; θ ) ) \theta_{map} = \operatorname*{argmin}_{\theta} - \frac{1}{N}\sum_n^N \log p(y_n| f(x_n;\theta)) θ ma p = θ argmin − N 1 n ∑ N log p ( y n ∣ f ( x n ; θ )) KL-Divergence (Forward) ¶ D KL [ p ∗ ( x ) ∣ ∣ p ( x ; θ ) ] = E x ∼ p ∗ [ log p ∗ ( x ) p ( x ; θ ) ] \text{D}_{\text{KL}}\left[ p_*(x) || p(x;\theta) \right] = \mathbb{E}_{x\sim p_*}\left[ \log \frac{p_*(x)}{p(x;\theta)}\right] D KL [ p ∗ ( x ) ∣∣ p ( x ; θ ) ] = E x ∼ p ∗ [ log p ( x ; θ ) p ∗ ( x ) ] This is the distance between the best distribution, p ∗ ( x ) p_*(x) p ∗ ( x ) p ( x ; θ ) p(x;\theta) p ( x ; θ )
There is an equivalence between the (Forward) KL-Divergence and the Maximum Likelihood Estimation. Maximizing the likelihood expresses it as maximizing the likelihood of the data given our estimated distribution. Whereas the KL-divergence is a distance measure between the parameterized distribution and the “true” or “best” distribution of the real data. They are equivalent formulations but the MLE equations shows how this is a proxy for fitting the “real” data distribution to the estimated distribution function.
D KL [ p ∗ ( x ) ∣ ∣ p ( x ; θ ) ] = E x ∼ p ∗ [ log p ∗ ( x ) p ( x ; θ ) ] \text{D}_{\text{KL}}\left[ p_*(x) || p(x;\theta) \right] = \mathbb{E}_{x\sim p_*}\left[ \log \frac{p_*(x)}{p(x;\theta)}\right] D KL [ p ∗ ( x ) ∣∣ p ( x ; θ ) ] = E x ∼ p ∗ [ log p ( x ; θ ) p ∗ ( x ) ] We can expand this term via logs
E x ∼ p ∗ [ log p ∗ ( x ) p ( x ; θ ) ] = E x ∼ p ∗ [ log p ∗ ( x ) − log p ( x ; θ ) ] \mathbb{E}_{x\sim p_*}\left[ \log \frac{p_*(x)}{p(x;\theta)}\right] = \mathbb{E}_{x\sim p_*}\left[ \log p_*(x) - \log p(x;\theta) \right] E x ∼ p ∗ [ log p ( x ; θ ) p ∗ ( x ) ] = E x ∼ p ∗ [ log p ∗ ( x ) − log p ( x ; θ ) ] The first expectation, E x ∼ p ∗ [ p ∗ ( x ) ] \mathbb{E}_{x\sim p_*}[p_*(x)] E x ∼ p ∗ [ p ∗ ( x )] entropy term (i.e. the expected uncertainty in the data). This is a constant term because no matter how well we estimate this distribution via our parameterized representation, p ( x ; θ ) p(x;\theta) p ( x ; θ )
E x ∼ p ∗ [ log p ∗ ( x ) p ( x ; θ ) ] = − E x ∼ p ∗ [ log p ( x ; θ ) ] \mathbb{E}_{x\sim p_*}\left[ \log \frac{p_*(x)}{p(x;\theta)}\right] = -\mathbb{E}_{x\sim p_*}\left[ \log p(x;\theta)\right] E x ∼ p ∗ [ log p ( x ; θ ) p ∗ ( x ) ] = − E x ∼ p ∗ [ log p ( x ; θ ) ] We can rewrite this in its integral form:
− E x ∼ p ∗ [ log p ( x ; θ ) ] = − ∫ log p ( x ; θ ) p ∗ ( x ) d x -\mathbb{E}_{x\sim p_*}\left[ \log p(x;\theta)\right] = - \int \log p(x;\theta) p_*(x)dx − E x ∼ p ∗ [ log p ( x ; θ ) ] = − ∫ log p ( x ; θ ) p ∗ ( x ) d x We will assume that the data distribution is a delta function, p ∗ ( x ) = δ ( x − x i ) p_*(x) = \delta (x - x_i) p ∗ ( x ) = δ ( x − x i )
− ∫ log p ( x ; θ ) p ∗ ( x ) d x = − ∫ log p ( x ; θ ) δ ( x − x i ) d x -\int \log p(x;\theta) p_*(x)dx = - \int \log p(x;\theta) \delta (x - x_i)dx − ∫ log p ( x ; θ ) p ∗ ( x ) d x = − ∫ log p ( x ; θ ) δ ( x − x i ) d x We will do the same approximation of the integral with samples from our delta distribution.
− ∫ log p ( x ; θ ) δ ( x − x i ) d x = − 1 N ∑ n N log p ( x n ; θ ) -\int \log p(x;\theta) \delta (x - x_i)dx = - \frac{1}{N}\sum_n^N \log p(x_n;\theta) − ∫ log p ( x ; θ ) δ ( x − x i ) d x = − N 1 n ∑ N log p ( x n ; θ ) So we have:
D KL [ p ∗ ( x ) ∣ ∣ p ( x ; θ ) ] = − 1 N ∑ n N log p ( x n ; θ ) = L N L L ( θ ) \text{D}_{\text{KL}}\left[ p_*(x) || p(x;\theta) \right] = - \frac{1}{N}\sum_n^N \log p(x_n;\theta) = \mathcal{L}_{NLL}(\theta) D KL [ p ∗ ( x ) ∣∣ p ( x ; θ ) ] = − N 1 n ∑ N log p ( x n ; θ ) = L N LL ( θ ) which exactly the function for the NLL Loss
Laplace Approximation ¶ This is where we approximate the posterior with a Gaussian distribution N ( μ , A − 1 ) \mathcal{N}(\mu, A^{-1}) N ( μ , A − 1 )
w = w m a p w=w_{map} w = w ma p p ( w ∣ D ) p(w|D) p ( w ∣ D )
A = ∇ ∇ log p ( D ∣ w ) p ( w ) A = \nabla\nabla \log p(D|w) p(w) A = ∇∇ log p ( D ∣ w ) p ( w )
Only captures a single mode and discards the probability mass
Sources
Variational Inference ¶ Definition : We can find the best approximation within a given family w.r.t. KL-Divergence.
KLD [ q ∣ ∣ p ] = ∫ w q ( w ) log q ( w ) p ( w ∣ D ) d w \text{KLD}[q||p] = \int_w q(w) \log \frac{q(w)}{p(w|D)}dw KLD [ q ∣∣ p ] = ∫ w q ( w ) log p ( w ∣ D ) q ( w ) d w Let q ( w ) = N ( μ , S ) q(w)=\mathcal{N}(\mu, S) q ( w ) = N ( μ , S ) ( q ∣ ∣ p ) (q||p) ( q ∣∣ p ) μ , S \mu, S μ , S
“Approximate the posterior, not the model” - James Hensman.
We write out the marginal log-likelihood term for our observations, y y y
log p ( y ; θ ) = E x ∼ p ( x ∣ y ; θ ) [ log p ( y ∣ θ ) ] \log p(y;\theta) = \mathbb{E}_{x \sim p(x|y;\theta)}\left[ \log p(y|\theta) \right] log p ( y ; θ ) = E x ∼ p ( x ∣ y ; θ ) [ log p ( y ∣ θ ) ] We can expand this term using Bayes rule: p ( y ) = p ( x , y ) p ( x ∣ y ) p(y) = p(x,y)p(x|y) p ( y ) = p ( x , y ) p ( x ∣ y )
log p ( y ; θ ) = E x ∼ p ( x ∣ y ; θ ) [ log p ( x , y ; θ ) ⏟ p r i o r − log p ( x ∣ y ; θ ) ⏟ p o s t e r i o r ] \log p(y;\theta) = \mathbb{E}_{x \sim p(x|y;\theta)}\left[ \log \underbrace{p(x,y;\theta)}_{prior} - \log \underbrace{p(x|y;\theta)}_{posterior}\right] log p ( y ; θ ) = E x ∼ p ( x ∣ y ; θ ) ⎣ ⎡ log p r i or p ( x , y ; θ ) − log p os t er i or p ( x ∣ y ; θ ) ⎦ ⎤ where p ( x , y ; θ ) p(x,y;\theta) p ( x , y ; θ ) p ( x ∣ y ; θ ) p(x|y;\theta) p ( x ∣ y ; θ )
We can use a variational distribution, q ( x ∣ y ; ϕ ) q(x|y;\phi) q ( x ∣ y ; ϕ )
log p ( y ; θ ) ≥ L E L B O ( θ , ϕ ) \log p(y;\theta) \geq \mathcal{L}_{ELBO}(\theta,\phi) log p ( y ; θ ) ≥ L E L BO ( θ , ϕ ) where L E L B O \mathcal{L}_{ELBO} L E L BO
L E L B O ( θ , ϕ ) = E q ( x ∣ y ; ϕ ) [ log p ( x , y ; θ ) − log q ( x ∣ y ; ϕ ) ] \mathcal{L}_{ELBO}(\theta,\phi) = \mathbb{E}_{q(x|y;\phi)}\left[ \log p(x,y;\theta) - \log q(x|y;\phi) \right] L E L BO ( θ , ϕ ) = E q ( x ∣ y ; ϕ ) [ log p ( x , y ; θ ) − log q ( x ∣ y ; ϕ ) ] we can rewrite this to single out the expectations. This will result in two important quantities.
L E L B O ( θ , ϕ ) = E q ( x ∣ y ; ϕ ) [ log p ( x , y ; θ ) ] ⏟ Reconstruction − D KL [ log q ( x ∣ y ; ϕ ) ∣ ∣ p ( x ; θ ) ] ⏟ Regularization \mathcal{L}_{ELBO}(\theta,\phi) = \underbrace{\mathbb{E}_{q(x|y;\phi)}\left[ \log p(x,y;\theta)\right]}_{\text{Reconstruction}} - \underbrace{\text{D}_{\text{KL}}\left[ \log q(x|y;\phi) || p(x;\theta)\right]}_{\text{Regularization}} L E L BO ( θ , ϕ ) = Reconstruction E q ( x ∣ y ; ϕ ) [ log p ( x , y ; θ ) ] − Regularization D KL [ log q ( x ∣ y ; ϕ ) ∣∣ p ( x ; θ ) ] Sampling Methods ¶ Monte Carlo ¶ We can produce samples from the exact posterior by defining a specific Monte Carlo chain.
We actually do this in practice with NNs because of the stochastic training regimes. We modify the SGD algorithm to define a scalable MCMC sampler.
Here
Markov Chain Monte Carlo ¶ Hamiltonian Monte Carlo ¶ Stochastic Gradient Langevin Dynamics ¶