Semantic Bottlenecks: Quantifying and Improving Inspectability of Deep Representations View Full Text


Ontology type: schema:Chapter     


Chapter Info

DATE

2021-03-17

AUTHORS

Max Maria Losch , Mario Fritz , Bernt Schiele

ABSTRACT

Today’s deep learning systems deliver high performance based on end-to-end training but are notoriously hard to inspect. We argue that there are at least two reasons making inspectability challenging: (i) representations are distributed across hundreds of channels and (ii) a unifying metric quantifying inspectability is lacking. In this paper, we address both issues by proposing Semantic Bottlenecks (SB), integrated into pretrained networks, to align channel outputs with individual visual concepts and introduce the model agnostic AUiC metric to measure the alignment. We present a case study on semantic segmentation to demonstrate that SBs improve the AUiC up to four-fold over regular network outputs. We explore two types of SB-layers in this work: while concept-supervised SB-layers (SSB) offer the greatest inspectability, we show that the second type, unsupervised SBs (USB), can match the SSBs by producing one-hot encodings. Importantly, for both SB types, we can recover state of the art segmentation performance despite a drastic dimensionality reduction from 1000s of non aligned channels to 10s of semantics-aligned channels that all downstream results are based on. More... »

PAGES

15-29

Identifiers

URI

http://scigraph.springernature.com/pub.10.1007/978-3-030-71278-5_2

DOI

http://dx.doi.org/10.1007/978-3-030-71278-5_2

DIMENSIONS

https://app.dimensions.ai/details/publication/pub.1136447494


Indexing Status Check whether this publication has been indexed by Scopus and Web Of Science using the SN Indexing Status Tool
Incoming Citations Browse incoming citations for this publication using opencitations.net

JSON-LD is the canonical representation for SciGraph data.

TIP: You can open this SciGraph record using an external JSON-LD service: JSON-LD Playground Google SDTT

[
  {
    "@context": "https://springernature.github.io/scigraph/jsonld/sgcontext.json", 
    "about": [
      {
        "id": "http://purl.org/au-research/vocabulary/anzsrc-for/2008/08", 
        "inDefinedTermSet": "http://purl.org/au-research/vocabulary/anzsrc-for/2008/", 
        "name": "Information and Computing Sciences", 
        "type": "DefinedTerm"
      }, 
      {
        "id": "http://purl.org/au-research/vocabulary/anzsrc-for/2008/0801", 
        "inDefinedTermSet": "http://purl.org/au-research/vocabulary/anzsrc-for/2008/", 
        "name": "Artificial Intelligence and Image Processing", 
        "type": "DefinedTerm"
      }
    ], 
    "author": [
      {
        "affiliation": {
          "alternateName": "Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbr\u00fccken, Germany", 
          "id": "http://www.grid.ac/institutes/grid.419528.3", 
          "name": [
            "Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbr\u00fccken, Germany"
          ], 
          "type": "Organization"
        }, 
        "familyName": "Losch", 
        "givenName": "Max Maria", 
        "id": "sg:person.013504324635.08", 
        "sameAs": [
          "https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.013504324635.08"
        ], 
        "type": "Person"
      }, 
      {
        "affiliation": {
          "alternateName": "CISPA Helmholtz Center for Information Security, Saarland Informatics Campus, Saarbr\u00fccken, Germany", 
          "id": "http://www.grid.ac/institutes/grid.507511.7", 
          "name": [
            "CISPA Helmholtz Center for Information Security, Saarland Informatics Campus, Saarbr\u00fccken, Germany"
          ], 
          "type": "Organization"
        }, 
        "familyName": "Fritz", 
        "givenName": "Mario", 
        "id": "sg:person.013361072755.17", 
        "sameAs": [
          "https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.013361072755.17"
        ], 
        "type": "Person"
      }, 
      {
        "affiliation": {
          "alternateName": "Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbr\u00fccken, Germany", 
          "id": "http://www.grid.ac/institutes/grid.419528.3", 
          "name": [
            "Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbr\u00fccken, Germany"
          ], 
          "type": "Organization"
        }, 
        "familyName": "Schiele", 
        "givenName": "Bernt", 
        "id": "sg:person.01174260421.90", 
        "sameAs": [
          "https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.01174260421.90"
        ], 
        "type": "Person"
      }
    ], 
    "datePublished": "2021-03-17", 
    "datePublishedReg": "2021-03-17", 
    "description": "Today\u2019s deep learning systems deliver high performance based on end-to-end training but are notoriously hard to inspect. We argue that there are at least two reasons making inspectability challenging: (i) representations are distributed across hundreds of channels and (ii) a unifying metric quantifying inspectability is lacking. In this paper, we address both issues by proposing Semantic Bottlenecks (SB), integrated into pretrained networks, to align channel outputs with individual visual concepts and introduce the model agnostic AUiC metric to measure the alignment. We present a case study on semantic segmentation to demonstrate that SBs improve the AUiC up\u00a0to four-fold over regular network outputs. We explore two types of SB-layers in this work: while concept-supervised SB-layers (SSB) offer the greatest inspectability, we show that the second type, unsupervised SBs (USB), can match the SSBs by producing one-hot encodings. Importantly, for both SB types, we can recover state of the art segmentation performance despite a drastic dimensionality reduction from 1000s of non aligned channels to 10s of semantics-aligned channels that all downstream results are based on.", 
    "editor": [
      {
        "familyName": "Akata", 
        "givenName": "Zeynep", 
        "type": "Person"
      }, 
      {
        "familyName": "Geiger", 
        "givenName": "Andreas", 
        "type": "Person"
      }, 
      {
        "familyName": "Sattler", 
        "givenName": "Torsten", 
        "type": "Person"
      }
    ], 
    "genre": "chapter", 
    "id": "sg:pub.10.1007/978-3-030-71278-5_2", 
    "isAccessibleForFree": false, 
    "isPartOf": {
      "isbn": [
        "978-3-030-71277-8", 
        "978-3-030-71278-5"
      ], 
      "name": "Pattern Recognition", 
      "type": "Book"
    }, 
    "keywords": [
      "deep learning system", 
      "semantic bottleneck", 
      "art segmentation performance", 
      "one-hot encoding", 
      "hundreds of channels", 
      "drastic dimensionality reduction", 
      "semantic segmentation", 
      "deep representation", 
      "end training", 
      "visual concepts", 
      "segmentation performance", 
      "learning system", 
      "network output", 
      "dimensionality reduction", 
      "high performance", 
      "channel output", 
      "inspectability", 
      "representation", 
      "segmentation", 
      "case study", 
      "network", 
      "performance", 
      "bottleneck", 
      "encoding", 
      "metrics", 
      "downstream results", 
      "output", 
      "channels", 
      "system", 
      "issues", 
      "concept", 
      "training", 
      "hundreds", 
      "alignment", 
      "work", 
      "SB type", 
      "types", 
      "quantifying", 
      "end", 
      "state", 
      "results", 
      "reasons", 
      "second type", 
      "AUIC", 
      "reduction", 
      "study", 
      "Sb layer", 
      "SSB", 
      "paper"
    ], 
    "name": "Semantic Bottlenecks: Quantifying and Improving Inspectability of Deep Representations", 
    "pagination": "15-29", 
    "productId": [
      {
        "name": "dimensions_id", 
        "type": "PropertyValue", 
        "value": [
          "pub.1136447494"
        ]
      }, 
      {
        "name": "doi", 
        "type": "PropertyValue", 
        "value": [
          "10.1007/978-3-030-71278-5_2"
        ]
      }
    ], 
    "publisher": {
      "name": "Springer Nature", 
      "type": "Organisation"
    }, 
    "sameAs": [
      "https://doi.org/10.1007/978-3-030-71278-5_2", 
      "https://app.dimensions.ai/details/publication/pub.1136447494"
    ], 
    "sdDataset": "chapters", 
    "sdDatePublished": "2022-12-01T06:54", 
    "sdLicense": "https://scigraph.springernature.com/explorer/license/", 
    "sdPublisher": {
      "name": "Springer Nature - SN SciGraph project", 
      "type": "Organization"
    }, 
    "sdSource": "s3://com-springernature-scigraph/baseset/20221201/entities/gbq_results/chapter/chapter_47.jsonl", 
    "type": "Chapter", 
    "url": "https://doi.org/10.1007/978-3-030-71278-5_2"
  }
]
 

Download the RDF metadata as:  json-ld nt turtle xml License info

HOW TO GET THIS DATA PROGRAMMATICALLY:

JSON-LD is a popular format for linked data which is fully compatible with JSON.

curl -H 'Accept: application/ld+json' 'https://scigraph.springernature.com/pub.10.1007/978-3-030-71278-5_2'

N-Triples is a line-based linked data format ideal for batch operations.

curl -H 'Accept: application/n-triples' 'https://scigraph.springernature.com/pub.10.1007/978-3-030-71278-5_2'

Turtle is a human-readable linked data format.

curl -H 'Accept: text/turtle' 'https://scigraph.springernature.com/pub.10.1007/978-3-030-71278-5_2'

RDF/XML is a standard XML format for linked data.

curl -H 'Accept: application/rdf+xml' 'https://scigraph.springernature.com/pub.10.1007/978-3-030-71278-5_2'


 

This table displays all metadata directly associated to this object as RDF triples.

135 TRIPLES      22 PREDICATES      73 URIs      66 LITERALS      7 BLANK NODES

Subject Predicate Object
1 sg:pub.10.1007/978-3-030-71278-5_2 schema:about anzsrc-for:08
2 anzsrc-for:0801
3 schema:author N3b66adb1c3874fefb6e8c689523cc669
4 schema:datePublished 2021-03-17
5 schema:datePublishedReg 2021-03-17
6 schema:description Today’s deep learning systems deliver high performance based on end-to-end training but are notoriously hard to inspect. We argue that there are at least two reasons making inspectability challenging: (i) representations are distributed across hundreds of channels and (ii) a unifying metric quantifying inspectability is lacking. In this paper, we address both issues by proposing Semantic Bottlenecks (SB), integrated into pretrained networks, to align channel outputs with individual visual concepts and introduce the model agnostic AUiC metric to measure the alignment. We present a case study on semantic segmentation to demonstrate that SBs improve the AUiC up to four-fold over regular network outputs. We explore two types of SB-layers in this work: while concept-supervised SB-layers (SSB) offer the greatest inspectability, we show that the second type, unsupervised SBs (USB), can match the SSBs by producing one-hot encodings. Importantly, for both SB types, we can recover state of the art segmentation performance despite a drastic dimensionality reduction from 1000s of non aligned channels to 10s of semantics-aligned channels that all downstream results are based on.
7 schema:editor N3be03a5b455b47bf96a3f2817df3dd65
8 schema:genre chapter
9 schema:isAccessibleForFree false
10 schema:isPartOf N8c64e281cffb41e5b9d2f88bec9abb0e
11 schema:keywords AUIC
12 SB type
13 SSB
14 Sb layer
15 alignment
16 art segmentation performance
17 bottleneck
18 case study
19 channel output
20 channels
21 concept
22 deep learning system
23 deep representation
24 dimensionality reduction
25 downstream results
26 drastic dimensionality reduction
27 encoding
28 end
29 end training
30 high performance
31 hundreds
32 hundreds of channels
33 inspectability
34 issues
35 learning system
36 metrics
37 network
38 network output
39 one-hot encoding
40 output
41 paper
42 performance
43 quantifying
44 reasons
45 reduction
46 representation
47 results
48 second type
49 segmentation
50 segmentation performance
51 semantic bottleneck
52 semantic segmentation
53 state
54 study
55 system
56 training
57 types
58 visual concepts
59 work
60 schema:name Semantic Bottlenecks: Quantifying and Improving Inspectability of Deep Representations
61 schema:pagination 15-29
62 schema:productId N402d0a9d69dc4b7f9c75bf525e46ca41
63 Na85f44cfe3164664b067c8d5a123da01
64 schema:publisher Nfceba5c1b5e94384aa2ff294391b3a7d
65 schema:sameAs https://app.dimensions.ai/details/publication/pub.1136447494
66 https://doi.org/10.1007/978-3-030-71278-5_2
67 schema:sdDatePublished 2022-12-01T06:54
68 schema:sdLicense https://scigraph.springernature.com/explorer/license/
69 schema:sdPublisher N40bf5842c6c642d7acb45da4cc94d4ca
70 schema:url https://doi.org/10.1007/978-3-030-71278-5_2
71 sgo:license sg:explorer/license/
72 sgo:sdDataset chapters
73 rdf:type schema:Chapter
74 N2bff8aa616d54ca8b51618d12cb32490 rdf:first sg:person.01174260421.90
75 rdf:rest rdf:nil
76 N2c990ce724794de9b81dd9c059501d1b rdf:first sg:person.013361072755.17
77 rdf:rest N2bff8aa616d54ca8b51618d12cb32490
78 N3b66adb1c3874fefb6e8c689523cc669 rdf:first sg:person.013504324635.08
79 rdf:rest N2c990ce724794de9b81dd9c059501d1b
80 N3be03a5b455b47bf96a3f2817df3dd65 rdf:first N3d2eada682b54ff1aa5f3b6c5111ba38
81 rdf:rest N73331706fac64d1d880c2f01301e4c2f
82 N3d2eada682b54ff1aa5f3b6c5111ba38 schema:familyName Akata
83 schema:givenName Zeynep
84 rdf:type schema:Person
85 N402d0a9d69dc4b7f9c75bf525e46ca41 schema:name doi
86 schema:value 10.1007/978-3-030-71278-5_2
87 rdf:type schema:PropertyValue
88 N40bf5842c6c642d7acb45da4cc94d4ca schema:name Springer Nature - SN SciGraph project
89 rdf:type schema:Organization
90 N73331706fac64d1d880c2f01301e4c2f rdf:first Naf9c810b77a3462e894d7fc14fac7773
91 rdf:rest N9dfa63a798b048bfb18e1acd5e020157
92 N8c64e281cffb41e5b9d2f88bec9abb0e schema:isbn 978-3-030-71277-8
93 978-3-030-71278-5
94 schema:name Pattern Recognition
95 rdf:type schema:Book
96 N9dfa63a798b048bfb18e1acd5e020157 rdf:first Nbd53b38e536143aeb4cdfcda70b1f9ec
97 rdf:rest rdf:nil
98 Na85f44cfe3164664b067c8d5a123da01 schema:name dimensions_id
99 schema:value pub.1136447494
100 rdf:type schema:PropertyValue
101 Naf9c810b77a3462e894d7fc14fac7773 schema:familyName Geiger
102 schema:givenName Andreas
103 rdf:type schema:Person
104 Nbd53b38e536143aeb4cdfcda70b1f9ec schema:familyName Sattler
105 schema:givenName Torsten
106 rdf:type schema:Person
107 Nfceba5c1b5e94384aa2ff294391b3a7d schema:name Springer Nature
108 rdf:type schema:Organisation
109 anzsrc-for:08 schema:inDefinedTermSet anzsrc-for:
110 schema:name Information and Computing Sciences
111 rdf:type schema:DefinedTerm
112 anzsrc-for:0801 schema:inDefinedTermSet anzsrc-for:
113 schema:name Artificial Intelligence and Image Processing
114 rdf:type schema:DefinedTerm
115 sg:person.01174260421.90 schema:affiliation grid-institutes:grid.419528.3
116 schema:familyName Schiele
117 schema:givenName Bernt
118 schema:sameAs https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.01174260421.90
119 rdf:type schema:Person
120 sg:person.013361072755.17 schema:affiliation grid-institutes:grid.507511.7
121 schema:familyName Fritz
122 schema:givenName Mario
123 schema:sameAs https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.013361072755.17
124 rdf:type schema:Person
125 sg:person.013504324635.08 schema:affiliation grid-institutes:grid.419528.3
126 schema:familyName Losch
127 schema:givenName Max Maria
128 schema:sameAs https://app.dimensions.ai/discover/publication?and_facet_researcher=ur.013504324635.08
129 rdf:type schema:Person
130 grid-institutes:grid.419528.3 schema:alternateName Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany
131 schema:name Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany
132 rdf:type schema:Organization
133 grid-institutes:grid.507511.7 schema:alternateName CISPA Helmholtz Center for Information Security, Saarland Informatics Campus, Saarbrücken, Germany
134 schema:name CISPA Helmholtz Center for Information Security, Saarland Informatics Campus, Saarbrücken, Germany
135 rdf:type schema:Organization
 




Preview window. Press ESC to close (or click here)


...