Keras
Francois Chollet released a wrapper in which a network is described in a dozen lines and the lower-level complexity is hidden.
Why it matters
Deep learning became accessible without understanding computation graphs, and the number of people who could train a network grew by an order of magnitude.
The first version ran on Theano, TensorFlow support followed, and in 2017 Keras became the latter's official high-level interface. Simplicity cost flexibility, and research groups mostly stayed at the lower level. But for applied work and teaching it became the usual entry point.