[bitL]

It's difficult to say to be honest; for me the "enthusiasm" works best, I simply picked an area I wanted to know (e.g. self-driving cars using DL) and then learned some mindblowing approaches, like NVidia/Tesla using a few layers of simple convolutional neural network and static images to predict steering angles, and then some people stacked RNN on top of this CNN and made it estimate steering angle from 10 previous frames and a current frame. See e.g. selfdrivingcars.mit.edu

If you are into CV, first start with very simple static image recognition with AlexNet/VGG/Inception etc. in Keras, try to understand CNNs a bit (it's inspired by biological neurons, they can do simple things like direction detection, edge detection etc. and overlap each other's field of vision; if you look at computational photography, convolutions do something similar, so the idea is why not use a layer of multiple convolutions, then make a hierarchy of those convolutional layers, and let the optimization/learning part of Deep Learning during training figure out what exact convolutions does it need instead of force-feeding them by hand). Play with the ways to improve training (batch normalization, image augmentation etc.) Once you understand this, your mind would probably explode and then it's time to understand RNNs/LSTMs/GANs and have fun applying it on voice, natural language, generating art etc.

You'll have a blast for sure when you realize what you can now easily do! Have fun! ;-)