Gaussian-Bernoulli RBM

Quote from the Paper of *To be Bernoulli or to be Gaussian, for a Restricted
Boltzmann Machine* 1

The Guassian-Bernoulli RGM(GBRBM) has visible units with real value $v_m$ and binary hidden units $h_n$. Based on the same idea as GBRGM, the energy function of GBRGM is defined as:

$$E(v,h) = -\sum_{i=1}^m\sum_{j=1}^n w_{ij}h_j\frac{v_i}{\sigma_i} - \sum_{i=1}^m\frac{(v_i-b_i)^2}{2\sigma_i^2} - \sum_{j=1}^mc_jh_j$$

Where $b_i$ and $c_i$ are biases corresponding respectively to visible and hidden units, $w_{ij}$ are the connecting weights between the visible and hidden units and $\sigma_i$ is the standard deviation associated with Guassian visible units $v_i$. Conditional probablities for visible and hidden units are:

$$p(v_i = v|h) = N(v|b_i + \sum_j h_jw_{ij}, \sigma^2)$$
$$p(h_j=1|v) = sigmoid(c_j + \sum w_{ij}\frac{v_i}{\sigma^2}$$

Where $N(\mu, \sigma^2)$ demotes the Gaussian probability destiny function with mean $\mu$ and standard deviation $\sigma$.

---

1. To be Bernoulli or to be Gaussian, for a Restricted
   Boltzmann Machine ↩