Things can get a little confusing when you first dive into learning about deep neural networks.

It helps to understand some basic ideas and rules of thumb.

The first is to think of problems as either classification or regression.

The second is to understand the problem domain. Are we dealing with images, language, objects described by a bunch of attributes, time series.

If the problem relates to classification of images then you should be considering convolutional neural networks.

You can lookup AlexNet and ResNet as examples of high performance architectures as inspiration.

If you want to create a model which processes objects described by a set of static attributes (say for heart disease diagnosis) then a simple feed forward dense network is a good starting point.

If you are looking to model problems which involves sequences (eg, text) or time services then you need to look at recurrent neural networks and their variants. Eg. LSTM.

At some point as your awareness develops, an understanding of ensemble methods and autoencoders is also helpful.

There are million other things to consider but this thread provides a pathway to help you navigate through what looks like a complex field.

A beginners guide to where to start and what you need to know to get going.