There are many different fields in DL and an extensive research in each one of them. The main 3 topics are computer vision that mainly use CNN architectures, NLP that mainly use RNN architectures (but not only, recent papers also combine CNNs in many NLP tasks) and general ML topics that use all the combinations of FC(fully connected layers ), CNN and RNN.
it depends on the nature of dataset. most popular and promising deep learning algorithms are sequential, CNN, RNN, LSTM, etc. CNN is more suitable for image based classification.
Bayesian deep learning over probabilistic/statistical models are realistic and more accurate techniques, however., such models are under consideration for the future data analysis development.
Deep learning is a huge sector. New algorithms are developed on a regular basis. Dropconnect is one of the method that gave state of the art result for MNIST dataset. It is not guaranteed that it will work excellent like this for other datasets.
There are many different fields in DL and an extensive research in each one of them. The main 3 topics are computer vision that mainly use CNN architectures, NLP that mainly use RNN architectures (but not only, recent papers also combine CNNs in many NLP tasks) and general ML topics that use all the combinations of FC(fully connected layers ), CNN and RNN.