In short, this architecture freezes the parameters and pathways used for previously learned tasks, and can learn new parameters and use new pathways for new tasks, with each new task learned faster than previous ones by leveraging all previously learned parameters and pathways (more efficient transfer learning).

It's a general neural net architecture.

Very cool.