Deep Belief Network's reasoning

20th-October-2020

Marginal distribution 𝑝𝑝 (𝐡𝐡 3 | 𝐯𝐯) can not be calculated analytically We introduce approximate distribution , propagate from input layer to hidden layer 𝐡𝐡 2 | 𝐡𝐡 3

Learning of DBN (pre-training)

Train sequentially using RBM • Learning by pre-training always increases log-likelihood