What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis

<p>Deep neural networks are inherently opaque and challenging to interpret. Unlike hand-crafted feature-based models, we struggle to comprehend the concepts learned and how they interact within these models. This understanding is crucial not only for debugging purposes but also for ensuring fa...

وصف كامل

محفوظ في:
التفاصيل البيبلوغرافية
المؤلف الرئيسي: Shammur Absar Chowdhury (16812394) (author)
مؤلفون آخرون: Nadir Durrani (5297438) (author), Ahmed Ali (705582) (author)
منشور في: 2024
الموضوعات:
الوسوم: إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
_version_ 1864513562842497024
author Shammur Absar Chowdhury (16812394)
author2 Nadir Durrani (5297438)
Ahmed Ali (705582)
author2_role author
author
author_facet Shammur Absar Chowdhury (16812394)
Nadir Durrani (5297438)
Ahmed Ali (705582)
author_role author
dc.creator.none.fl_str_mv Shammur Absar Chowdhury (16812394)
Nadir Durrani (5297438)
Ahmed Ali (705582)
dc.date.none.fl_str_mv 2024-01-01T00:00:00Z
dc.identifier.none.fl_str_mv 10.1016/j.csl.2023.101539
dc.relation.none.fl_str_mv https://figshare.com/articles/journal_contribution/What_do_end-to-end_speech_models_learn_about_speaker_language_and_channel_information_A_layer-wise_and_neuron-level_analysis/23918445
dc.rights.none.fl_str_mv CC BY 4.0
info:eu-repo/semantics/openAccess
dc.subject.none.fl_str_mv Information and computing sciences
Artificial intelligence
Human-centred computing
Language, communication and culture
Linguistics
Speech
Neuron-level analysis
Interpretability
Diagnostic classifier
AI explainability
End-to-end architecture
dc.title.none.fl_str_mv What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
dc.type.none.fl_str_mv Text
Journal contribution
info:eu-repo/semantics/publishedVersion
text
contribution to journal
description <p>Deep neural networks are inherently opaque and challenging to interpret. Unlike hand-crafted feature-based models, we struggle to comprehend the concepts learned and how they interact within these models. This understanding is crucial not only for debugging purposes but also for ensuring fairness in ethical decision-making. In our study, we conduct a post-hoc functional interpretability analysis of pretrained speech models using the probing framework (Hupkes et al., 2018). Specifically, we analyze utterance-level representations of speech models trained for various tasks such as speaker recognition and dialect identification. We conduct layer and neuron-wise analyses, probing for speaker, language, and channel properties. Our study aims to answer the following questions: (i) what information is captured within the representations? (ii) how is it represented and distributed? and (iii) can we identify a minimal subset of the network that possesses this information? Our results reveal several novel findings, including: (i) channel and gender information are distributed across the network, (ii) the information is redundantly available in neurons with respect to a task, (iii) complex properties such as dialectal information are encoded only in the task-oriented pretrained network, (iv) and is localized in the upper layers, (v) we can extract a minimal subset of neurons encoding the pre-defined property, (vi) salient neurons are sometimes shared between properties, (vii) our analysis highlights the presence of biases (for example gender) in the network. Our cross-architectural comparison indicates that: (i) the pretrained models capture speaker-invariant information, and (ii) CNN models are competitive with Transformer models in encoding various understudied properties.</p><h2>Other Information</h2><p>Published in: Computer Speech & Language<br>License: <a href="http://creativecommons.org/licenses/by/4.0/" rel="noreferrer" target="_blank">http://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1016/j.csl.2023.101539" target="_blank">https://dx.doi.org/10.1016/j.csl.2023.101539</a></p>
eu_rights_str_mv openAccess
id Manara2_950e4b3545ce1e973d14f7c610457cbf
identifier_str_mv 10.1016/j.csl.2023.101539
network_acronym_str Manara2
network_name_str Manara2
oai_identifier_str oai:figshare.com:article/23918445
publishDate 2024
repository.mail.fl_str_mv
repository.name.fl_str_mv
repository_id_str
rights_invalid_str_mv CC BY 4.0
spelling What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysisShammur Absar Chowdhury (16812394)Nadir Durrani (5297438)Ahmed Ali (705582)Information and computing sciencesArtificial intelligenceHuman-centred computingLanguage, communication and cultureLinguisticsSpeechNeuron-level analysisInterpretabilityDiagnostic classifierAI explainabilityEnd-to-end architecture<p>Deep neural networks are inherently opaque and challenging to interpret. Unlike hand-crafted feature-based models, we struggle to comprehend the concepts learned and how they interact within these models. This understanding is crucial not only for debugging purposes but also for ensuring fairness in ethical decision-making. In our study, we conduct a post-hoc functional interpretability analysis of pretrained speech models using the probing framework (Hupkes et al., 2018). Specifically, we analyze utterance-level representations of speech models trained for various tasks such as speaker recognition and dialect identification. We conduct layer and neuron-wise analyses, probing for speaker, language, and channel properties. Our study aims to answer the following questions: (i) what information is captured within the representations? (ii) how is it represented and distributed? and (iii) can we identify a minimal subset of the network that possesses this information? Our results reveal several novel findings, including: (i) channel and gender information are distributed across the network, (ii) the information is redundantly available in neurons with respect to a task, (iii) complex properties such as dialectal information are encoded only in the task-oriented pretrained network, (iv) and is localized in the upper layers, (v) we can extract a minimal subset of neurons encoding the pre-defined property, (vi) salient neurons are sometimes shared between properties, (vii) our analysis highlights the presence of biases (for example gender) in the network. Our cross-architectural comparison indicates that: (i) the pretrained models capture speaker-invariant information, and (ii) CNN models are competitive with Transformer models in encoding various understudied properties.</p><h2>Other Information</h2><p>Published in: Computer Speech & Language<br>License: <a href="http://creativecommons.org/licenses/by/4.0/" rel="noreferrer" target="_blank">http://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1016/j.csl.2023.101539" target="_blank">https://dx.doi.org/10.1016/j.csl.2023.101539</a></p>2024-01-01T00:00:00ZTextJournal contributioninfo:eu-repo/semantics/publishedVersiontextcontribution to journal10.1016/j.csl.2023.101539https://figshare.com/articles/journal_contribution/What_do_end-to-end_speech_models_learn_about_speaker_language_and_channel_information_A_layer-wise_and_neuron-level_analysis/23918445CC BY 4.0info:eu-repo/semantics/openAccessoai:figshare.com:article/239184452024-01-01T00:00:00Z
spellingShingle What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
Shammur Absar Chowdhury (16812394)
Information and computing sciences
Artificial intelligence
Human-centred computing
Language, communication and culture
Linguistics
Speech
Neuron-level analysis
Interpretability
Diagnostic classifier
AI explainability
End-to-end architecture
status_str publishedVersion
title What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
title_full What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
title_fullStr What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
title_full_unstemmed What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
title_short What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
title_sort What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis
topic Information and computing sciences
Artificial intelligence
Human-centred computing
Language, communication and culture
Linguistics
Speech
Neuron-level analysis
Interpretability
Diagnostic classifier
AI explainability
End-to-end architecture