interface

Interfaces enable modularity. In general, standardizing an interface can yield quadratic benefit at linear cost.

Suppose we have $M$ people and $N$ cars, but that every car had its own unique control scheme. Then for every person to learn to drive every car would require separate training for each person-car pair, or $O(MN)$ total effort. Adopting a common interface brings the work down to $O(M + N)$, since each car and each person just have to adapt themselves to that interface.

In deep learning, shared representations can function as a common interface for models.