The world is being quietly rearranged by people who write very long documents.


The title they went with Massive Redundancy in Gradient Transport Enables Sparse Online Learning Noisy translates that to

Recurrent neural networks can learn online with 94% fewer computations


Researchers discovered that recurrent neural networks can learn in real time using only 6% of the gradient calculations normally required, without losing accuracy. This means AI systems that need to adapt instantly to changing conditions can do so much faster and cheaper, especially as they get larger.
Online learning — where a system adapts to new information in real time rather than waiting for a batch of training data — has been computationally expensive enough to be impractical; this makes it feasible at scale by showing the redundancy in gradient computation was far larger than anyone assumed.

If you insist
Read the original →