Available toolboxes are making things easy. you may use deep learning as a toolbox but with not an adequate insight into the concepts. Better to start with shallow networks as they still have satisfactory performance and good enough in many applications.
It is essential to learn firstly basic concepts of Machine Learning; those two approaches; supervised and unsupervised, some techniques of pre-processing of data, features selection algorithms, be familiar with the datasets and then go to deep learning.
As a formal way of learning, starting with the basics and progressively advancing is a proper way of learning (not just machine learning). But this should not prevent you from looking on the other side. If you have the basics you can quickly find different resources that simplify the advanced part of deep learning.
I say this because sometimes you may have a problem that's requires application of deep learning, hence if you want to follow the normal learning curve it may take much time which is usually not in our control.