ebook img

Transparency and Interpretability for Learned Representations of Artificial Neural Networks PDF

230 Pages·2022·5.313 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Transparency and Interpretability for Learned Representations of Artificial Neural Networks

Richard Meyes Transparency and Interpretability for Learned Representations of Artificial Neural Networks Transparency and Interpretability for Learned Representations of Artificial Neural Networks Richard Meyes Transparency and Interpretability for Learned Representations of Artificial Neural Networks Dr.-Ing.RichardMeyes,M.Sc. ResearchGroupLead“InterpretableLearningModels” InstituteforTechnologiesandManagementofDigital Transformation UniversityofWuppertal Wuppertal,Germany Thescientificstudiesandcorrespondingresultspresentedinthisbookwereconducted withintheframeworkofadoctoralprojecttoobtainthedegreeofDr.-Ing.attheSchool ofElectrical,InformationandMediaEngineeringattheUniversityofWuppertal. ISBN978-3-658-40003-3 ISBN978-3-658-40004-0 (eBook) https://doi.org/10.1007/978-3-658-40004-0 ©TheEditor(s)(ifapplicable)andTheAuthor(s),underexclusivelicensetoSpringer FachmedienWiesbadenGmbH,partofSpringerNature2022 Thisworkissubjecttocopyright.AllrightsaresolelyandexclusivelylicensedbythePublisher, whetherthewholeorpartofthematerialisconcerned,specificallytherightsoftranslation,reprint- ing, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physicalway,andtransmissionorinformationstorageandretrieval,electronicadaptation,computer software,orbysimilarordissimilarmethodologynowknownorhereafterdeveloped. Theuseofgeneraldescriptivenames,registerednames,trademarks,servicemarks,etc.inthis publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexempt fromtherelevantprotectivelawsandregulationsandthereforefreeforgeneraluse. Thepublisher,theauthors,andtheeditorsaresafetoassumethattheadviceandinformationinthis bookarebelievedtobetrueandaccurateatthedateofpublication.Neitherthepublishernorthe authorsortheeditorsgiveawarranty,expressedorimplied,withrespecttothematerialcontained hereinorforanyerrorsoromissionsthatmayhavebeenmade.Thepublisherremainsneutralwith regardtojurisdictionalclaimsinpublishedmapsandinstitutionalaffiliations. This Springer Vieweg imprint is published by the registered company Springer Fachmedien WiesbadenGmbH,partofSpringerNature. Theregisteredcompanyaddressis:Abraham-Lincoln-Str.46,65189Wiesbaden,Germany Menoughttoknowthatfromthebrain, andfromthebrainonly,ariseour pleasures,joys,laughter,andjests,as wellasoursorrows,pains,griefs,and tears.Throughit,inparticular,wethink, see,hear,anddistinguishtheuglyfrom thebeautiful,thebadfromthegood,the pleasantfromtheunpleasant,insome casesusingcustomasatest,inothers perceivingthemfromtheirutility.Itisthe samethingwhichmakesusmador delirious,inspiresuswithdreadorfear, whetherbynightorbyday,brings sleeplessness,inopportunemistakes, aimlessanxieties,absent-mindedness, andactsthatarecontrarytohabit.These thingsthatwesufferallcomefromthe brain,whenitisnothealthy,butbecomes abnormallyhot,cold,moist,ordry,or suffersanyotherunnatural affectionto whichitwasnotaccustomed.Madness comesfromitsmoistness.Whenthebrain isabnormallymoist,ofnecessityit moves,andwhenitmovesneithersight norhearingarestill,butweseeorhear nowonethingandnowanother,andthe tonguespeaksinaccordancewiththe thingsseenandheardonanyoccasion. Butallthetimethebrainisstillamanis intelligent. —Hippocrates TheSacredDisease,inHippocrates, trans.W.H.S.Jones(1923), Vol.2,175 Contents 1 Introduction .................................................. 1 1.1 Object of Investigation .................................... 2 1.2 Research Questions ....................................... 3 1.3 Structure of the Thesis .................................... 6 2 Background&Foundations .................................... 9 2.1 A Short History of AI Research ............................ 9 2.1.1 The Early Years .................................... 10 2.1.2 The Golden Ages ................................... 11 2.1.3 The AI Winter ..................................... 13 2.1.4 The AI Renaissance ................................ 13 2.2 The Modern Era of AI Research ............................ 14 2.2.1 Deep Learning for Computer Vision .................. 15 2.2.2 Computer Vision Beyond the ILSVRC ................ 17 2.2.3 From Supervised Learning to Reinforcement Learning .......................................... 18 2.2.4 Deep Reinforcement Learning Breakthroughs for Video and Board Games ......................... 19 2.2.5 Tackling Games with Imperfect Information ........... 23 2.3 Towards Research on Transparency & Interpretability ......... 24 2.3.1 A Shift of Paradigm: From Optimizing to Understanding ................................... 25 2.3.2 Inspirations from Neuroscience Research .............. 27 3 MethodsandTerminology ..................................... 31 3.1 Learned Representations ................................... 31 3.1.1 Investigating Learned Representations ................. 33 vii viii Contents 3.1.1.1 Statistical Measures ........................ 34 3.1.1.2 Transformations & Embeddings .............. 35 3.1.2 Visualizing Structure of the Learned Representations ... 46 3.1.2.1 Magnitude of Neuron Activations ............ 47 3.1.2.2 Selectivity of Neuron Activations ............. 47 3.1.2.3 Ablations of Individual Neurons .............. 48 3.1.2.4 Gini Importance ............................ 51 3.2 Delimitation of the Object of Investigation ................... 51 3.2.1 Transparency for Computer Vision Models ............ 52 3.2.2 Transparency for Motion Control Models .............. 53 3.2.3 Transfer to an Industrial Application Scenario .......... 55 4 RelatedWork ................................................. 57 4.1 Relationship Between a Network’s Input Features and its Output ................................................... 59 4.2 Visualization of Network Properties and Graphical User Interfaces ................................................ 61 4.3 Investigating the Importance of Individual Network Components .............................................. 64 4.3.1 Miscellaneous Contributions ......................... 65 4.3.2 Ablations Studies ................................... 67 4.3.3 Reverse Engineering of Neural Networks .............. 69 5 ResearchStudies .............................................. 71 5.1 Investigating Learned Representations in Computer Vision ..... 75 5.1.1 Research Study 1: Characterizing Single Neurons in a Shallow MLP .................................. 75 5.1.1.1 Key Contributions of the Study .............. 75 5.1.1.2 Methods and Experimental Design ........... 77 5.1.1.3 Results .................................... 78 5.1.1.4 Summary and Contribution of the Results to the Research Questions ................... 88 5.1.2 Research Study 2: Network Ablations in a Deep Neural Network .................................... 91 5.1.2.1 Key Contributions of the Study .............. 91 5.1.2.2 Methods and Experimental Design ........... 92 5.1.2.3 Results .................................... 93 5.1.2.4 Summary and Contribution of the Results to the Research Questions ................... 98 Contents ix 5.1.3 Research Study 3: Functional Neuron Populations in Custom-made CNNs ............................. 100 5.1.3.1 Key Contributions of the Study .............. 100 5.1.3.2 Methods and Experimental Design ........... 101 5.1.3.3 Results .................................... 105 5.1.3.4 Summary and Contribution of the Results to the Research Questions ................... 113 5.2 Investigating Learned Representations in Motor Control ....... 114 5.2.1 Research Study 4: Influence of Network Ablations on Activation Patterns ............................... 115 5.2.1.1 Key Contributions of the Study .............. 116 5.2.1.2 Methods and Experimental Design ........... 117 5.2.1.3 Results .................................... 119 5.2.1.4 Summary and Contribution of the Results to the Research Questions ................... 128 5.2.2 Research Study 5: Relation Between Neural Activations and Agent Behavior ...................... 129 5.2.2.1 Key Contributions of the Study .............. 130 5.2.2.2 Methods and Experimental Design ........... 131 5.2.2.3 Results .................................... 135 5.2.2.4 Summary and Contribution of the Results to the Research Questions ................... 147 6 TransferStudies .............................................. 149 6.1 Transfer Study 1: Network Ablations for Deep Drawing ....... 150 6.1.1 Key Contributions of the Study ...................... 151 6.1.2 Methods and Experimental Design .................... 151 6.1.3 Results ............................................ 159 6.1.4 Summary and Contribution of the Results to the Research Questions ........................... 164 6.2 Transfer Study 2: Attention Mechanisms for Deep Drawing .... 166 6.2.1 Key Contributions of the Study ...................... 167 6.2.2 Methods and Experimental Design .................... 167 6.2.3 Results ............................................ 173 6.2.4 Summary and Contribution of the Results to the Research Questions ........................... 181 x Contents 7 CriticalReflection&Outlook .................................. 183 7.1 Reflection of Results & Contribution to Research Questions .... 183 7.1.1 Research Question 1 ................................ 184 7.1.2 Research Question 2 ................................ 185 7.1.3 Research Question 3 ................................ 187 7.1.4 Research Question 4 ................................ 189 7.2 Future Research Directions ................................. 191 8 Summary ..................................................... 195 References ....................................................... 197

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.