Reversible designs for extreme memory cost reduction of cnn training

HIGHLIGHTS

• who: Tristan Hascoet from the Kobe University, u2011, Kobe, u20110013, Japan have published the paper: Reversible designs for extreme memory cost reduction of CNN training, in the Journal: (JOURNAL)
• what: The authors investigate the propagation of numerical errors in long chains of invertible operations and analyze their effect on training. The authors introduce the notion of pixel-wise memory cost to characterize the memory footprint of model training and propose a new model architecture able to efficiently train arbitrarily deep neural networks with a minimum memory cost of 352 bytes per input pixel. For instance . . .
  

   

  

  If you want to have access to all the content you need to log in!

  Thanks :)

  Username or E-mail:

  If you don't have an account, you can create one here.