tutorials
play

Tutorials Interpretable Deep Learning: Towards Understanding & - PowerPoint PPT Presentation

Tutorials Interpretable Deep Learning: Towards Understanding & Explaining DNNs P a r t 3 : V a l i d a t i n g E x p l a n a t i o n s W o j c i e c h S a m e k , G r g o i r e M o n t


  1. Tutorials Interpretable Deep Learning: Towards Understanding & Explaining DNNs P a r t 3 : V a l i d a t i n g E x p l a n a t i o n s W o j c i e c h S a m e k , G r é g o i r e M o n t a v o n , K l a u s - R o b e r t M ü l l e r 1 / 2 8

  2. F r o m M L S u c c e s s e s t o A p p l i c a t i o n s Deep Net outperforms Medical Diagnosis humans in image classification AlphaGo beats Go human champ Autonomous Driving Visual Reasoning Networks (smart grids, etc.) 2 / 2 8

  3. M a k i n g M L M o d e l s I n t e r p r e t a b l e 3 / 2 8

  4. L a y e r - W i s e R e l e v a n c e P r o p a g a t i o n ( L R P ) [ B a c h ’ 1 5 ] 4 / 2 8

  5. D e e p T a y l o r D e c o m p o s i t i o n [ M o n t a v o n ’ 1 7 ] S u p p o s e t h a t w e h a v e p r o p a g a t e d t h e r e l e v a n c e u n t i l Q u e s t i o n : a g i v e n l a y e r . H o w s h o u l d i t b e p r o p a g a t e d o n e l a y e r f u r t h e r ? I d e a : B y p e r f o r m i n g a T a y l o r e x p a n s i o n o f t h e r e l e v a n c e . 5 / 2 8

  6. D e e p T a y l o r D e c o m p o s i t i o n R e l e v a n c e n e u r o n : T a y l o r e x p a n s i o n : R e d i s t r i b u t i o n : 6 / 2 8

  7. R e v i s i t i n g t h e D T D R o o t P o i n t ( D e e p T a y l o r g e n e r i c ) C h o i c e o f r o o t p o i n t 1 . n e a r e s t r o o t ✔ 2 . r e s c a l e d e x c i t a t i o n s ✔ 3 . g e n e r a l i z e d G e n e r a l i z e d r u l e 7 / 2 8

  8. T h e S p e c i a l C a s e “ = 1 . 0 ” γ F i n d t h e d i f f e r e n c e . . . I s t h e r e a c o n n e c t i o n b e t w e e n t h e t w o m e t h o d s ? Q u e s t i o n : 8 / 2 8

  9. T h e S p e c i a l C a s e “ = 1 . 0 ” γ w h i c h c a n a l s o b e r e w r i t t e n a s : [Shrikumar’17] Not Just a F o r n e t w o r k s w i t h b i a s z e r o , t h e Black Box: Learning p r o c e d u r e b e c o m e s e q u i v a l e n t t o Important Features Through Propagating Activation 9 / 2 8 g r a d x i n p u t [ s e e a l s o S h r i k u m a r ’ 1 7 ] Differences, ArXiv

  10. Question: How to select the optimal parameter “ γ ” ? 1 0 / 2 8

  11. E x p l a n a t i o n S e l e c t i o n 1 1 / 2 8

  12. E x p l a n a t i o n S e l e c t i o n M o r e d i r e c t a p p r o a c h : T r y a l l p a r a m e t e r s , a n d s e l e c t t h e o n e p r o d u c i n g t h e b e s t e x p l a n a t i o n s . H o w t o a s s e s s e x p l a n a t i o n q u a l i t y ? Q u e s t i o n : 1 2 / 2 8

  13. E v a l u a t i n g E x p l a n a t i o n s H u m a n a s s e s s m e n t - A e s t h e t i c p r o p e r t i e s - U s a b i l i t y o f t h e e x p l a n a t i o n ( e . g . t o u n d e r s t a n d t h e c l a s s i fi e r ) . → R e q u i r e s a n e x p e r i me n t a l s t u d y . 1 3 / 2 8

  14. E v a l u a t i n g E x p l a n a t i o n s I d e a : T e s t i n g i f e x p l a n a t i o n s s a t i s f y c e r t a i n a x i o m s / p r o p e r t i e s . E x a mp l e s : - E x p l a n a t i o n m u s t b e s e l f - c o n s i s t e n t ( e . g . ) c o n s e r v a t i o n o f e v i d e n c e - E x p l a n a t i o n m u s t b e c o n s i s t e n t i n i n p u t d o m a i n ( e . g . ) c o n t i n u i t y - E x p l a n a t i o n m u s t b e c o n s i s t e n t i n t h e s p a c e o f m o d e l s ( e . g . i m p l e m e n t a t i o n i n v a r i a n c e ) 1 4 / 2 8

  15. E x a m p l e 1 : C o n s e r v a t i o n S i m p l e e x a m p l e : P o s s i b l e e x p l a n a t i o n s : 1 5 / 2 8

  16. E x a m p l e 1 : C o n s e r v a t i o n 1 6 / 2 8

  17. E x a m p l e 1 : C o n s e r v a t i o n 1 7 / 2 8

  18. W h y G r a d x I n p u t S c o r e s E x p l o d e ? A n s w e r : y d e p t h 1 0 1 1 2 N e u r a l n e t w o r k d e p t h c a u s e s x y t h e f u n c t i o n t o b e c o m e s t e e p 2 -1 x 0 a n d t h e g r a d i e n t v e r y l a r g e . 0 1 -1 y [ c f . B e n g i o ’ 9 4 , M o n t u f a r ’ 1 4 ] 0 0 d e p t h 2 1 1 1 2 2 x y 2 -1 2 -1 x [Bengio’94] Learning long- 0 0 1 -1 -1 term dependencies with gradient descent is difficult. y 0 0 0 IEEE Trans. Neural Networks d e p t h 3 1 1 1 1 2 2 2 x y [Montufar’14] On the Number 2 -1 2 -1 2 -1 x of Linear Regions of DNNs. 0 1 8 / 2 8 0 1 -1 -1 -1 NIPS 2014.

  19. W h y G r a d x I n p u t S c o r e s E x p l o d e ? T h i s c a n a l s o b e s e e n f r o m t h e f o r m u l a s : d i v i s i o n b y z e r o 1 9 / 2 8

  20. E x a m p l e 2 : C o n t i n u i t y [ M o n t a v o n ’ 1 8 ] E x p l a n a t i o n s c o r e s m u s t b e c o n t i n u o u s i n i n p u t d o m a i n . 2 0 / 2 8

  21. C o n t i n u i t y D e m o v i d e o i n p u t A n i m a t i o n s a v a i l a b l e a t : 2 1 / 2 8 h t t p : / / w w w . h e a t m a p p i n g . o r g / e v a l u a t i n g

  22. W h y i s G r a d x I n p u t D i s c o n t i n u o u s ? A n s w e r : y d e p t h 1 0 1 1 2 A g a i n , b e c a u s e o f d e p t h , x y s p e c i fi c a l l y , b e c a u s e t h e 2 -1 x 0 f u n c t i o n b e c o m e s h i g h l y 0 1 -1 n o n - s m o o t h . y 0 0 d e p t h 2 1 1 1 2 2 [ c f . M o n t u f a r ’ 1 4 , B a l d u z z i ’ 1 7 ] x y 2 -1 2 -1 x 0 [Montufar’14] On the Number 0 1 -1 -1 of Linear Regions of DNNs. y 0 0 0 NIPS 2014. d e p t h 3 1 1 1 1 2 2 2 x y [Balduzzi’17] The Shattered Gradients Problem: If resnets 2 -1 2 -1 2 -1 x 0 2 2 / 2 8 are the answer [...] ICML 2017 0 1 -1 -1 -1

  23. E x a m p l e 3 : I m p l . I n v a r i a n c e [ S u n d a r a r a j a n ’ 1 7 ] t w o n e t w o r k s i m p l e m e n t i n g t h e m a x i m u m f u n c t i o n : E x a mp l e : C o u n t e r - e x a m p l e f o r : G r a d i e n t i s i m p l e m e n t a t i o n i n v a r i a n t , t h e r e f o r e e x p l a n a t i o n t o o . N e t w o r k ( a ) : N e t w o r k ( b ) : [Sundararajan’17] M Sundararajan, A Taly, Q Yan: Axiomatic Attribution for Deep 2 3 / 2 8 Networks. ICML 2017

  24. I m p l e m e n t a t i o n M a t t e r s f o r L R P n a i v e i m p l e m e n t a t i o n s b e t t e r i m p l e m e n t a t i o n 2 4 / 2 8

  25. A B l i n d S p o t i n E x p l a n a t i o n S e l e c t i o n C o n s i d e r t h e s i m p l e e x p l a n a t i o n t e c h n i q u e : r e d i s t r i b u t i n g u n i f o r m l y o n p i x e l s . I t i s : - c o n s e r v a t i v e - c o n t i n u o u s - i m p l e m e n t a t i o n i n v a r i a n t i t i s a l s o c o m p l e t e l y u n i n f o r m a t i v e . b u t → N e e d t o v e r i f y ( i . e . t h e e x p l a n a t i o n s h o u l d s e l e c t i v i t y d i s c r i m i n a t e b e t w e e n r e l e v a n t a n d i r r e l e v a n t v a r i a b l e s . ) 2 5 / 2 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