· 5 2 n 4; q r • ~2013.3, phd@: -, 9 • 2013.4~2016.3, -/ @ibm$ • 2016.4~2017.8, -/ @)...

47
1

Upload: others

Post on 21-May-2020

7 views

Category:

Documents


0 download

TRANSCRIPT

����������������

� �

���� ���� �

1

�������

5�2�

n4;��Q��R• ~2013.3, PhD@:�-, 9�• 2013.4~2016.3, -/�@IBM$� .-• 2016.4~2017.8, -/�@)%ERATO, NII• 2017.9~, � @:�-, 9�

n-/• *�8�

� DLHBCMJGM='7�3 QECML’11R� ,��!=� QAISTATS’15,17R

• (&�3JGM=6"� ANENIM#=0+� QAISTATS’18R� JGM��<?@KPF=1���� QAAAI’17,18R� ���O6"�>=”��” QICML’19R

2

����

n YouTube�����

3

��'�

n����&"#( ���• ���%��• ���'��

n ���• � �% �

• ���% �

n��%�!��&�#$• *)+'��

4

AI/����0/1,��%��

n � *')• �#�(+��2��

n ��*')• ��2��+����+��

5

����%��(&*#" !-+�� ���+��

�����

��

��

�����

!.�$+�#

AI/����<;=5��,��n� 3.2

• #)�14��>�

n��3.2• ��>� 4����4��

6

�����

��

XX�����

XX�����

���

����,��1-9( &'64�"!�4��� %$��4��

.:38�/0�*9+7

[=�\li0�-g8?

n�*-h$"�4i=��_D8:aqe]pZ

• ��� OS�&)��!�3�.JIMKLN�� C��i��

• OSz��~���u��k�}�~�)5jYOS|~���jOSz��~i� i#<�6��k�2 i=��6�h+�cpZ

� sxv�����t�=�A��i��• AIz��~���u��k�}�~�)5jY'B5-�)5�kF��)5r�m~��y���h�bsxv�����tr dco^�npZ

• QW� �7����@9��RPTU�� $"�4r�nY��{hGcp��%�h=�A�r>cZ

• WV� =��6OS�XOSJIPOUTO��}wy��� [�F_(;b�Hf`pOS\i/1E,Z

7

���� ���������� ������

8

�� EU��� ������GDPR�n GDPR-221. The data subject shall have the right not to be subject to a decision based

solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her.

2. Paragraph 1 shall not apply if the decision: is necessary for entering into, or performance of, a contract between the data subject and a data controller; is authorised by Union or Member State law to which the controller is subject and which also lays down suitable measures to safeguard the data subject’s rights and freedoms and legitimate interests; or is based on the data subject’s explicit consent.

3. In the cases referred to in points (a) and (c) of paragraph 2, the data controller shall implement suitable measures to safeguard the data subject’s rights and freedoms and legitimate interests, at least the right to obtain human intervention on the part of the controller, to express his or her point of view and to contest the decision.

4. Decisions referred to in paragraph 2 shall not be based on special categories of personal data referred to in Article 9(2)1), unless point (a) or (g) of Article 9(2) applies and suitable measures to safeguard the data subject’s rights and freedoms and legitimate interests are in place.

9

W�AX�=��\I(�El�F�kG�

n ��9B�E�5�J� P��• U��9B�E�5�DH4�+�Vl�kURg]qAIVmlt���~l�ehbfI(�E\7pLnrfYqT

n @12,8• U/����9B������*�EO6VlK%l�kSAIl���l jim�^dt���~l�ehbfI(�E\"_orfYqT

n <��E#C.-�JST�• :>O6�5{�}�[oU��5����x��AI�3yz��l����R�s;cq&��|��uvt��l;?V\�OardT`l�gI(�E\N

F�EQ�l�ehbf�p�_orfYqT

n ')@1��M��

• UAI03�5Vl�gAIl���w��wz�kG�bfZpSI(�El:>O6l�F�\!$arfYqT 10

"�@�

n�?2016��02���#/,@1�*;%-�V(��?/9D)�6��• ICML, NIPS>=@���#@MKP�*;C%-�V(��@TWHJRKP6.�8E:4D3

n��'A<B+�• ���#?57D%-�X!@OKHQWHY, � �$, Vol.33,

No.3, pages 366--369, 2018.• LGWPSWNUI@����F�%9D��, Qiita&�

11

�����

n �AI����� �������

12

Peeking inside the black-box: A survey on Explainable Ar>ficial Intelligence (XAI)hCps://ieeexplore.ieee.org/document/8466590/

7� ������SCOPUS, IEEExplore, ACM Digital Library, Google Scholar, Citeseer Library, ScienceDirect, arXiv����"%��&�������� “intelligible”,

“interpretable”, “transparency”, “black box”, “understandable”, “comprehensible”,

“explainable”���AI�& ��“Ar>ficial Intelligence”, “Intelligent system”, “Machine learning”, “deep

learning”, “classifier” , “decision tree”���# ���'$ !����

��

��'�

n����&"#( ���• ���%��• ���'��

n ���• � �% �

• ���% �

n��%�!��&�#$• *)+'��

13

LA�MI

n�J�J,0S�E'�@R1�#H2*��KH:4• 7GTH*�=�&<8KVYUO��IPCE�HR4

� ��K3� "=5>TH*�9CBQ�?NH:Z6F:;�*I�D:E*��J� /�S-ME:R>F= :4

n�%�H*��• 1. .&H��J��• 2. .&H�!VYUJ��• 3. $�()IPR*�• 4. WVXJ�+• …

14

��:;�,��*.�&14�!2�"'�����+9�(5�8<�

Q. $4�3-#6������������������6��� =���

→/�)%��70

���'��),389/

1. ��'��• ��)��($*260-40-.517)��)� +��$*#&%��&$*��!

2. ���'��• ��'260-40-.517+���) "����'517%��$*#&%��&$*��!

15

������!

��"→ →�#�����$��"′→ →�#′

�����

�����

��������

16

��!��#���!��

n ��#��"��$+.)&,)&'-*/#��#��% ��$� ��� �$���

n ��#�� !����% ��$��• Why Should I Trust You?: Explaining the Predic<ons of Any

Classifier, KDD'16 [Python�� LIME; R�� LIME]

n ��#�� !��*0(% ��$��• Understanding Black-box Predic<ons via Influence Func<ons,

ICML’17 [Python�� influence-release]

17

LIME')*��n Why Should I Trust You?: Explaining the Predictions of

Any Classifier, KDD'16 [Python�� LIME; R�� LIME]• &(� ��'��#$"�+��!*�• .-/+���-0,(��%��.-/%��!*�

� ��.-/(��(�%��� (�����+�*�

18

LIME�����

19

LIME����n �������

n ,)-�)+(&• vs *'%.���• ������� ������$���

→ LIME� !�,)-��$��$�����"���#�"� 20

��!��#���!��

n ��#��"��$+.)&,)&'-*/#��#��% ��$� ��� �$���

n ��#�� !����% ��$��• Why Should I Trust You?: Explaining the Predictions of Any

Classifier, KDD'16 [Python�� LIME; R�� LIME]

n ��#�� !��*0(% ��$��• Understanding Black-box Predictions via Influence Functions,

ICML’17 [Python�� influence-release]

21

influence����n Understanding Black-box Predictions via Influence

Functions, ICML’17 [Python�� influence-release]

n ��"#!("′, %′)����������� �"#!"����������� ��$

22

� ����������

����������������

������������������

influence+� 35743

n �/7=6("′, %′)����'&�)#&.��7=6"(+��,*0 .��1/�>

n :7<2 % = (("; *+), *+2��"0&8;9=6)$/�*+ = argmin2∈4 5

67(8,9)∈:;(<; +)

*+=6> = argmin2∈4 56∈: ?@A 6B6>

;(<; +)

n <> = ("′, %′)+��( *+,*0 .���$/�>• C+=6> − C+ 2��#&��

� ��- *+=6> − *+ ���!0-<>+��,�)��%��

23

������� �

<> = ("′, %′)����� �

influence4����

n !" = (%′, (′)4��/ *+51:&7���)9"<• ,+-." − ,+ ;��(+��

� �!6 *+-." − *+ #�$':6!"4��5�0�!* �

n ��2��• �.4!" = (%′, (′)3-�. ,+-.";�)9�

• ����#""8)%9�

n influence• ,+-." − ,+;��� 4��;�,.���3��)9�

,+-." − ,+ ≈ − 12 345-678(!"; ,+)

24

������������

�����������

influence��n Data Poisoning

• ����� '+& ��������'+&���($%#����������'+&#�"�

• ������'+&�����)'*������� '+&����"!���"�

25

��/���12-3

n��1��0(4;>96<967=:?1��1��5��(4'-,��-(4��"

n��1��-/+)� �5��(4��• LIMEA .1��� &��0��*+)%��(4"

n��1��-/+):@85��(4��• influenceA� ��1��5�$'-,!�:@81� ��#5��(4

26

������.�����

n ��!"$),���• Interpretable Predictions of Tree-based Ensembles

via Actionable Feature Tweaking, KDD’17

• '&+��� ���"*%���������!"$),� (-#����� �

27

����� ��

❌ 30����������200��

20����������200��

30��������300��

��

�…

��������…��

��������

28

�����������

n ���%(#!&#!"'$) ��������� �'$)��������������

n Born Again Treesn Making Tree Ensembles Interpretable: A Bayesian Model

Selec;on Approach, AISTATS'18 [Python�� defragTrees]

29

BATrees$'(��

n Born Again Trees• ��#3/6*��&(3/6� 9�� :"���(�

n ��• ���)�3/6!*� !���/8,*��(�

� ��%57-2$+4.06�!��

� 516%3/6!*� !���

• ��)����/8,*� !�� *���(�

30

"# → →�$#�����!

… …

"% → →�$%

"& → →�$&"# → →�$#

������'

… …

"% → →�$%

"& → →�$&

defragTrees!$&��n Making Tree Ensembles Interpretable: A Bayesian Model

Selec;on Approach, AISTATS'18 [Python�� defragTrees]• �� )(*'��"*+*����&�• ����*+*#��!���

n�,�� �%��"�"��"����

31

���� whenRelationship ≠ Not-in-family, WifeCapital Gain < 7370

���� whenRelationship ≠ Not-in-familyCapital Gain >= 7370

���� whenRelationship ≠ Not-in-family, UnmarriedCapital Gain < 5095Capital Loss < 2114

���� whenRelationship = Not-in-familyCountry ≠ China, PeruCapital Gain < 5095

���� whenRelationship ≠ Not-in-familyCountry ≠ ChinaCapital Gain < 5095

���� whenRelationship ≠ Not-in-familyCapital Gain >= 7370

… …

defragTrees?CE��

n RNSG��@$�W��SVSX= �0��9E7>=RNS@��PTMLG��9E1

n ()*,'+&,))-@IQVKY

32

�$�5��SVS?��12��.5$�%?�;:D��

�/G!913

HUJUOS�A$�5��8F<4E1��@$�W��SVSXG�"9E@A #1

��8F:RNSG��@$�W��SVSX= �0��9FBRNSG�"=6E1

���

R

��'�

n����&"#( ���• ���%��• ���'��

n ���• � �% �

• ���% �

n��%�!��&�#$• *)+'��

33

���)(*'�

���&�� ��&�� ���

��

�� ���� ��������� "%%� $��� ���!#

����5@ED6 �%����!.*163%#

n ����16"����4�8;=<56����.*163%#

n ��16����CBD4:<��>�,0"BEA5��5��75�>�7<#• ����CBD4"BEA5��5��7>'0 +%#

35

��5��(1)<:&43/-:F

��

��� ����CBD��5��462?3��($<

5FG

�1)3%#19��61)<#

����5@ED6 �%����!.*163%#

n ����16"����4�8;=<56����.*163%#

n ��16����CBD4:<��>�,0"BEA5��5��75�>�7<#• ����CBD4"BEA5��5��7>'0 +%#

36

��5��(1)<:&43/-:F

��

��� ����CBD��5��462?3��($<

5FG

�1)3%#19��61)<#

�� ���%���"���'�����(#���&$����→ ��� �����!'�����

�;�079A��9��C��3A:<

�;�5 *�%�;',KHMC�-)+

• � KHMPLasso(EJOEKHLNDQ� ���.@(8;��/&�.C#=�BA)

• MOMKHMPMOMLEI(MOMFGI(���Q� If-ThenMOM6KHM/!$1BA4>�#�/',)� ��=C "�2?3,)

37

���7�$HFI5=�4:�$'L

n �6!� �8�+�@(1$HFIA�23.9%"

n � �8��&?"• � �6�8#>&5$"FJE7GCBD!HFI7GCBD�'=!��$8�,>&?"

38

��7��(4)?<%6521<K

�� � �HFI��7�68��+(/-*�$<KK

$;!0@8""

2�+I�@Q\Z^3;�ORH7G�Q5

n ZaY�FN=6Q�0• *.�J17&�\Z^M^a^\Z^S�DEL555

n"!4>JN8H��IKZaY�';�/�J$)S[T]`V?E4$)S��@QN8I-�\Z^S��?E7Q5

39

���;9:?75>J\Z^���c

U`W`J�I��@QKAJU`W` ,�J��;X_CB555

^a^I%�F<H75>J\Z^���c

(�;17G# IHPM@7KAHJI4^a^I(�;�E>H7b

1�'J�@PZX[2Q&D>PKL,?84

n IB,?8;]• �:�� �</=ENY\W\VIZX[L��@P4

• �+ K�&G7P�����IZX[FI8G�'@P��Q&�AI84

n���KSUXTS• 1"8ZX[2Q$�&D>E3�+ J��?COH9;]

→ 5ZX[Q$���@P.6Q!:P4

40

0G89��Q��?M@4)%I-�L4445(*6G5#6

9�R]

�+

3�)L�BR\Z]4S([email protected]:6

n KD.A:=_• �<��!�>1?GP[^Y^XK\Z]N �BR6

• �-�M�(I9R����K\Z]HK:I�)BR��S(�CK:6

n�� MUWZVU• 3$:\Z]4S&�(F@G5�-�L��AEQJ;=_

→ 7\Z]S&���BR08S#<R6

41

2I:;��S��AOB6+'K/�N6667*,8I7%�8

;�T_

7�+8I7%"�8

;�T_

�-�

3�)N�CT_]`4V([email protected]:6

n ME.B:=c• �<��!�>1?IR^a\a[M_]`P �CT6

• �-�O�(K9T����M_]`JM:K�)CT��V(�DM:6

n�� OXZ]YX• 3$:_]`4V&�(H@I5�-�N��BFSL;=c

→ 7_]`V&���CT08V#<T6

42

2K:;��V��BQC6+'M/�P6667*,8K7%�8

;�Wc

7�+8K7%"�8

;�Wc

7�+8K7%�8

AUGbb

�-�

���=(�

&�=,/JILF#�� 6D��F��6D.

n���=�54N P IMH�%=�)

• ���;�'�$=8@<-��;�&2#�BED.

n���=�54O P KMG= ���

• JIL�"F#���6D3:9-KMG2731BAC �=,/JILF*?D.

43

#+;JIL>339>!0;/.

��8�� + ��1. �*��

• ��,��+'-361/41/0527,��,��.��'-&)(��)'-��$

2. ���*��• �"*361/41/0527.� �,#%�!��*527(��'-&)(��)'-��$

44

� ���!

!&" → →�&#(���$!&"′ → →�&#′

�����%

�����%

�%�'�� →"������#��� �Ulrich et al., ICML’19�

�%�'�� →"�������� ���� $����

�$

n �����$ !&����• ���#��• ���%��

n ����• ���#��

• ���#��

n ���#���$�!"• (')%�

45

�/C��0��

n ���.��20��• �&:+!9���0�$1�@>B;����/��496"/)9%-,��;��)9�

• ���.��,1�@>B-��0�/<A=?#�(9�

• *38����.��1�'$.!��# 9�� ������0�/1��'$.!�50# 9%-#�7:+!9�

46

����������������������������

��2>��3��

n��1��� 24��3���/��&?• LIME/influence10/#��1��<��'� /(7&?

n�.&6),#3��3���4$7&?• �.&6),3!+*4?• �2��+-���3���4?

� �"�#9=8#:9;#��10257 %?

47