deep virtual networks for memory efficient inference of
play

Deep Virtual Networks for Memory Efficient Inference of Multiple - PowerPoint PPT Presentation

Deep Virtual Networks for Memory Efficient Inference of Multiple Tasks (CVPR 2019) Reading Group August 21, 2019 Computer Vision Lab @ ETH Zurich Suman Saha (postdoc) M o t t i i v v a a t i o n i t i s w e l l - k n


  1. Deep Virtual Networks for Memory Efficient Inference of Multiple Tasks (CVPR 2019) Reading Group August 21, 2019 Computer Vision Lab @ ETH Zurich Suman Saha (postdoc)

  2. M o t t i i v v a a t i o n  i t i s w e l l - k n o w n t h a t D N N s h a v e a h i i g g h m m e e m m o r y y f f o o o t t p p r i i n n t [ 1 0 , 1 7 ] l i m i t i n g t h e i r p r a c t i c a l a p p l i c a t i o n s , s u c h a s m o b i l e p h o n e s , r o b o t s , a n d a u t o n o m o u s v e h i c l e s o f l o w c a p a c i t y  T o a d d r e s s t h i s , r e s e a r c h a i m e d a t r e d u c i i n n g t t h h e n n u m b e e r r o o f f p p a r a a m m e e t t e e r r s [ 1 2 , 1 4 ] , h o w e v e r , t h e r e i s a t r a d e - o fg b e t w e e n a c c u r a c y a n d t h e n u m b e r o f p a r a m e t e r s ( m e m o r y b u d g e t ) a t t e s t t i m e  w e e w w a a n n t t a a n n e e t t w w o o r r k k t t h h a a t t g g i i v v e e s s c c o m p e e t t i i t t i i v e p p e e r r f o o r r m a a n n c e e u n d e e r r a a g g i v e n m e e m m o o r r y y b u d g e e t t  B e s i d e s , g i v e n N d i fg e r e n t m e m o r y b u d g e t s , w e d e fj n e a n d t r a i n N d i fg e r e n t D N N m o d e l s w h i c h r e q u i r e a d d i i t t i i o o n a l t t r r a i i n n i i n n g c c o o s s t 2

  3. M o t t i i v v a a t i o n  [ 1 9 , 2 2 ] p r o p o s e d s i n g l e D N N m o d e l s w h i c h c a n p e r f o r m m u l t t i i p p l e e i n f e e r r e e n n c e u n d e e r r d d i i fg fg e r e e n n t t m e e m m o o r r y y b b u d g e e t t s s a l l o w i n g fm e x i b l e a c c u r a c y - m e m o r y t r a d e - o fg s w i t h i n a s i i n n g l e e n n e t t w w o o r r k ( a l s o c a l l e d m e m o r y e ffj c i e n t i n f e r e n c e )  a n d t h u s , c a n a v o i d i n t r o d u c i n g m u l t i p l e n e t w o r k s f o r d i fg e r e n t m e m o r y b u d g e t ( n o t e , t h e s e a r e f o r a s i i n n g l e e t t a a s s k )  l e a r n i n g m u l t i p l e t a s k s s i m u l t a n e o u s l y i n a n e t w o r k a v o i d m u l t i - s t a g e t r a i n i n g [ 2 , 2 6 ] a n d i m p r o v e g e n e r a l i z a t i o n [ 5 , 7 , 3 9 ]  t h i s w o r k p r o p o s e s a n a p p r o a c h t h a t p e r f o r m s m e e m m o o r r y y e e ffj ffj c i i e e n n t t i n f e e r r e e n n c e f o r m u l t t i i p p l e e t t a a s k k s s [ [ * ] i n a s i i n n g l e e n n e e t t w w o o r r k [19] Eunwoo Kimet al. NestedNet : Learning nested sparse structures in deep neural networks. CVPR 2018 [22] Gustav Larsson et al. FractalNet : Ultra-deep neural networks without residuals. ICLR 2017 [*] Multiple tasks refer to multiple datasets, unless stated otherwise 3

  4. A A p p p r o a c h  a r c h i t e c t u r e c o n t a i n i n g m u l t t i i p p l e e n n e e t t w w o o r r k k s s o f d d i i fg fg e r e e n n t c c o o n n fj fj g u r a a t t i i o o n n s t e r m e d d e e e p p v i i r r t t u u a a l l n n e t t w w o o r r k k s s ( D V N s )  E a c h D V N s h a r e s p a r a m e t e r s o f t h e a r c h i t e c t u r e a n d p e r f o r m s m e m o r y e ffj c i e n t i n f e r e n c e f o r i t s c o r r e s p o n d i n g t a s k 4

  5. A A p p p r o a c h  t h e p r o p o s e d a r c h i t e c t u r e i s b a s e d o n a b a a c c k k b b o n e e n n e e t t w w o o r r k  t h e n e t w o r k p a r a m e t e r s a r e d i v i d e d i n t o m u l t i p l e d i s j o i n t u n i i t t s s  u n i i t t s a r e c o l l e c t e d b y d i v i d i n g  a s e t o f f e a t u r e m a p s i n e a c h  l a y e r i n t o m u l t i p l e s u b s e t s  A D V N i s s t r u c t u r e d h i e r a r c h i c a l l y  w h i c h c o n t a i n s m u l t i p l e l e v e l s  o f h i e r a r c h y c o r r e s p o n d i n g t o  d i fg e r e n t n u m b e r s o f u n i t s e n a b l i n g  m u l t i p l e i n f e r e n c e f o r d i fg e r e n t  m e m o r y b u d g e t s 5

  6. A A p p p r o a c h  a u n i t c a n b e s h a r e d b y d i fg e r e n t D V N s a l l o w i n g m u l t i p l e D V N s i n a s i n g l e d e e p n e t w o r k t o p e r f o r m m u l t t i i - - t t a a s s k k i i n n g  E a c h D V N h a s a u n i i q q u e e c c o o n n fj fj g u r a a t t i i o o n ( S e c t i o n 3 . 2 ) ( i . e . , a h i e r a r c h i c a l s t r u c t u r e w i t h a d i fg e r e n t o r d e r o f u n i t s ) ,  a n d i s s p e c i a a l l i i z z e d f f o o r r a a s s i i n n g l e e t t a a s k  Tie a p p r o a c h i s r e a l i z e d i n a s i i n n g l e  t t r r a i i n n i i n n g s s t t a a g e b a s e d o n a s i n g l e  b a c k b o n e a r c h i t e c t u r e  ( e . g . , a r e s i d u a l n e t w o r k [ 1 3 ] ) ,  w h i c h s i g n i fj c a n t l y r e d u c e e s s  t t r r a i i n n i i n n g e e fg fg o r t t s s a n d  n e e t t w w o o r r k k s s t t o o r r a g e 6

  7. A A p p p r o a c h – M M e m o r y e ffj ffj c c i i e n n t t l l e a r n n i i n n g g  g i v e n a b a c k b o n e , d i v i d e t h e n e t w o r k p a r a m e t e r s i n t o k d i s j o i n t s u b s e t s : i . e . W = [ W_ { 1 } , ../ , W_ { k } ]  l - t h l e v e l o f h i e r a r c h y ( l > = 2 ) c o n t a i n s t h e s u b s e t s i n t h e ( l - 1 ) - t h l e v e l a n d o n e a d d i t i o n a l s u b s e t  l e v e l - 1 ( l o w e s t l e v e l ) c o n t a i n s ( l = 1 )  a s i n g l e s u b s e t s  l e v e l n _ { h } ( h i g h e s t l e v e l ) c o n t a i n s  a l l s u b s e t s ( i . e . W)  n _ { h } : n u m b e r o f l e v e l s o f h i e r a r c h y  k : n u m b e r o f s u b s e t s  i n t h i s w o r k , k = n _ { h }  e a c h l e v e l o f h i e r a r c h y d e fj n e s  a n e t w o r k c o r r e s p o n d i n g t o t h e s u b s e t  a n d p r o d u c e s a n o u t p u t 7

  8. A A p p p r o a c h – M M e m o r y e ffj ffj c c i i e n n t t l l e a r n n i i n n g g  L o s s f u n c t i o n :  : d a t a s e t - i m a g e - l a b e l p a i r s  : l e v e l s o f h i e r a r c h y  : s e t o f p a r a m e t e r s a n d c a n b e o p t i m i z e d b y t h e s u m o f t h e l o s s f u n c t i o n s  w h e r e :  i s a s e t o f p a r a m e t e r s o f t h a t a r e a s s i g n e d t o t h e l - t h l e v e l a n d d e s i g n e d b y f o l l o w i n g g r o u p - w i s e p r u n i n g a p p r o a c h e s [ 1 4 , 3 3 ]  N o o t t e e , E E q q . ( ( 1 ) i s a a p p p l i i e e d t t o o a a s i i n n g l e e t t a a s s k 8

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend