Why are RBMs not popular?

Question

Neural networks

I got the impression that at conferences the lion's share of reports are about CNN, RNN and autoencoders. And they don’t say anything new about RBM. What happened?

Answer 1

I found the answer myself: because the main use case of RBM was the initialization of network weights through pre-training in order to eliminate the problem of gradient degradation. But since then, more efficient ways to solve this problem have been found. In particular, glorot initialization and ReLU.