ÂÅÐÑÈß ÄËß ÑËÀÁÎÂÈÄßÙÈÕ
Enter
Login:
Password:
Forgot your password?
scientific activity
structureeducational projectsperiodicalsstaffpress centercontacts
ðóññêèé | english

Structural Models and Deep Learning

Joint IITP RAS, NRU HSE &Skoltech seminar  

November 22 (Tuesday), 1830, IITP, room 615    
 
Alexander Novikov (HSE)
 
Tensorizing Neural Networks

Convolutional neural networks excel in image recognition tasks, but this comes at the cost of high computational and memory complexity. CNNs require millions of floating point operations to process an image and therefore real-time applications need powerful CPU or GPU devices. Moreover, these networks contain millions of trainable parameters and consume hundreds of megabytes of storage and memory bandwidth. Thus, CNNs are forced to use RAM instead of solely relying on the processor cache – orders of magnitude more energy efficient memory device – which increases the energy consumption even more. These reasons restrain the spread of CNNs on mobile devices. I will talk about our work on tensor factorization framework to compress fully-connected and convolutional layers of CNNs. Another research direction (besides compression) is to increase the size of the layers by training them in the compact tensor format to increase the accuracy. 

22.11.2016 |
 

 

© Institute for Information Transmission Problems of the Russian Academy of Sciences (Kharkevich Institute), 2024
About  |  Contacts  |  Ïðîòèâîäåéñòâèå êîððóïöèè