Deep learning is a big umbrella that covers many projects and neural network configurations using a variety of methods. The following chart can give you an idea of the variety of architectures available. There is typically a trade-off between the accuracy obtainable and required computing time to train the models that has to be considered. The right trade-off depends on the use case. The y axis represents accuracy and the x axis represents compute requirements:

This is a big field and there are many areas to explore. There are methods that specialize in feature extraction from images, sound, and text. There are methods for classifying thousands to millions of images. There are methods for word translation that remember some previous words already translated in a sentence to improve accuracy on the current word being translated.

Some use cases combine network architecture types. The following is an example of a use case that extracts features from an image, then uses a Convolutional Neural Net (CNN) to classify elements in the image, and uses a Recurrent Neural Net (RNN) to match the extracted features to word choice to create a text caption for the original image: