<?xml version="1.0" encoding="UTF-8"?><mets:mets xmlns:mets="http://www.loc.gov/METS/" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:tef="http://www.abes.fr/abes/documents/tef" xmlns:metsRights="http://cosimo.stanford.edu/sdr/metsrights/" xmlns:dcterms="http://purl.org/dc/terms/" xmlns:mads="http://www.loc.gov/mads/">
    <mets:metsHdr ID="rennes1-ori-wf-1-20080" CREATEDATE="2024-10-10T01:26:34" LASTMODDATE="2024-10-10T01:26:35">
  <mets:agent ROLE="CREATOR">
            <mets:name>Université de Rennes</mets:name>
        </mets:agent>
</mets:metsHdr>
    <mets:dmdSec ID="desc_expr" CREATED="2024-10-10T01:26:34">
  <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_desc_these">
            <mets:xmlData>
                <tef:thesisRecord>
     <dc:title xml:lang="en">Neural network quantization methods for FPGA on-board processing of satellite images</dc:title>
     <dcterms:alternative xml:lang="fr">Méthodes de quantification des réseaux de neurones pour le traitement embarqué d'images satellites sur FPGA</dcterms:alternative>
     <dc:subject xml:lang="fr">Réseaux de neurones profonds</dc:subject><dc:subject xml:lang="fr">Précision limitée</dc:subject><dc:subject xml:lang="fr">Quantification pendant l'entraînement</dc:subject><dc:subject xml:lang="fr">Optimisation adaptative de la largeur de bits</dc:subject><dc:subject xml:lang="fr">Petit flottant</dc:subject><dc:subject xml:lang="fr">Détection de navires</dc:subject><dc:subject xml:lang="fr">Segmentation sémantique</dc:subject><dc:subject xml:lang="fr">FPGA
</dc:subject>
     <dc:subject xml:lang="en">Deep neural networks</dc:subject><dc:subject xml:lang="en">Reduced precision</dc:subject><dc:subject xml:lang="en">Quantization aware training</dc:subject><dc:subject xml:lang="en">Adaptive bit-width optimization</dc:subject><dc:subject xml:lang="en">Minifloat</dc:subject><dc:subject xml:lang="en">Ship detection</dc:subject><dc:subject xml:lang="en">Semantic segmentation</dc:subject><dc:subject xml:lang="en">FPGA</dc:subject>
     <tef:sujetRameau><tef:vedetteRameauNomCommun>
						<tef:elementdEntree autoriteSource="Sudoc" autoriteExterne="030971098">Réseaux neuronaux (informatique)</tef:elementdEntree>
					</tef:vedetteRameauNomCommun><tef:vedetteRameauNomCommun>
						<tef:elementdEntree autoriteSource="Sudoc" autoriteExterne="223540633">Apprentissage profond</tef:elementdEntree>
					</tef:vedetteRameauNomCommun><tef:vedetteRameauNomCommun>
						<tef:elementdEntree autoriteSource="Sudoc" autoriteExterne="028025644">Imagerie satellitaire</tef:elementdEntree>
					</tef:vedetteRameauNomCommun><tef:vedetteRameauNomCommun>
						<tef:elementdEntree autoriteSource="Sudoc" autoriteExterne="034756957">Réseaux logiques programmables par l'utilisateur</tef:elementdEntree>
					</tef:vedetteRameauNomCommun></tef:sujetRameau>
     
     
     
     <dcterms:abstract xml:lang="fr">L'apprentissage profond commence à être employé avec succès dans diverses applications spatiales. Cependant, en raison de son empreinte mémoire et de son volume de calcul, la phase d'inférence de nombreux modèles est encore principalement réalisée sur des plateformes terrestres. Des recherches récentes se sont intéressées à la quantification des réseaux de neurones afin de réduire ce volume de calcul et de faciliter le traitement embarqué. Cette thèse étudie la quantification des réseaux de neurones et son impact sur la précision des prédictions des modèles destinés aux accélérateurs FPGA pour les applications spatiales embarquées. Nous introduisons une méthode d'optimisation pour la quantification de nombres entiers uniformes en précision mixte des poids et des activations. Sa caractéristique principale est l'utilisation de largeurs de bits fractionnaires relâchés mises à jour à l'aide d'une règle de descente de gradient, mais par ailleurs discrétisées pour toutes les opérations (lors les calculs progressif et rétrogressif). Nous présentons également une approche performante basée sur la quantification en virgule flottante de faible précision (appelée minifloats). Notre méthode adapte une approche d'entraînement de réseaux de neurones profonds quantifiés, principalement utilisée pour la quantification en nombres entiers/virgule fixe. Enfin, nous avons étudié l'impact des sauts de connexion longs des modèles de segmentation sémantique en forme de U sur la précision des prédictions et mis en évidence leur empreinte mémoire.</dcterms:abstract>
     <dcterms:abstract xml:lang="en">Deep learning is starting to be successfully applied to various space applications. However, due to its memory footprint and computational intensity, inference of many models is still mainly performed on ground platforms. Recent research has focused on neural network quantization to mitigate this computational burden and facilitate on-board processing. This thesis investigates neural network quantization and its impact on model accuracy targeting FPGA accelerators for on-board space applications. We introduce an optimization-based method for mixed-precision uniform quantization of both weights and activations. Its defining characteristic is the use of relaxed fractional bit-widths that are updated using a gradient descent rule but otherwise discretized for all operations (in forward and backward pass computations). We also present an efficient approach based on low-precision floating-point quantization (so-called minifloats). Our method adapts a quantized deep neural network training approach predominantly used for integer/fixed-point-based quantization. Finally, we studied the impact of long skip connections in U-shaped semantic segmentation models on accuracy and highlighted their memory footprint.</dcterms:abstract>
     <dc:type>Electronic Thesis or Dissertation</dc:type><dc:type xsi:type="dcterms:DCMIType">Text</dc:type>
     <dc:language xsi:type="dcterms:RFC3066">en</dc:language>
    </tef:thesisRecord>
            </mets:xmlData>
        </mets:mdWrap>
</mets:dmdSec>
    <mets:dmdSec ID="desc_edition" CREATED="2024-10-10T01:26:34">
  <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_desc_edition">
            <mets:xmlData>
                <tef:edition><dcterms:medium xsi:type="dcterms:IMT">application/pdf</dcterms:medium><dcterms:extent>1 : 13972 Ko</dcterms:extent><dc:identifier xsi:type="dcterms:URI">https://ged.univ-rennes1.fr/nuxeo/site/esupversions/7810be56-fee4-42fb-b300-32258d8ef1c6</dc:identifier></tef:edition>
            </mets:xmlData>
        </mets:mdWrap>
</mets:dmdSec>
    <mets:amdSec>
        <mets:techMD ID="admin_expr">
            <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_admin_these">
                <mets:xmlData>
                    <tef:thesisAdmin>
                        <tef:auteur>
       <tef:nom>Gernigon</tef:nom>
       <tef:prenom>Cédric</tef:prenom>
       
       <tef:dateNaissance>1996-03-16</tef:dateNaissance>
       <tef:nationalite scheme="ISO-3166-1">fr</tef:nationalite>
       <tef:autoriteExterne autoriteSource="Sudoc">285911481</tef:autoriteExterne>
       <tef:autoriteExterne autoriteSource="mailPerso">cedric.gernigon-contact@mailo.bzh	</tef:autoriteExterne>
      </tef:auteur>
                        <dc:identifier xsi:type="tef:NNT">2024URENS122</dc:identifier>
                        <dc:identifier xsi:type="tef:nationalThesisPID">http://www.theses.fr/2024URENS122</dc:identifier>
                        <dcterms:dateAccepted xsi:type="dcterms:W3CDTF">2024-12-18</dcterms:dateAccepted>
                        <tef:thesis.degree>
                            <tef:thesis.degree.discipline xml:lang="fr">Informatique</tef:thesis.degree.discipline>
                            <tef:thesis.degree.grantor>
        <tef:nom>Université de Rennes</tef:nom><tef:autoriteInterne>thesis.degree.grantor_1</tef:autoriteInterne>
        
        <tef:autoriteExterne autoriteSource="Sudoc">26693823X</tef:autoriteExterne>
       </tef:thesis.degree.grantor>
                            <tef:thesis.degree.level>Doctorat</tef:thesis.degree.level>
                        </tef:thesis.degree>
                        <tef:theseSurTravaux>non</tef:theseSurTravaux>
                        <tef:avisJury>oui</tef:avisJury><tef:directeurThese><tef:nom>Sentieys</tef:nom><tef:prenom>Olivier</tef:prenom><tef:autoriteInterne>intervenant_1</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">061585920</tef:autoriteExterne></tef:directeurThese><tef:presidentJury><tef:nom>Ménard</tef:nom><tef:prenom>Daniel</tef:prenom><tef:autoriteInterne>intervenant_2</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">069860505</tef:autoriteExterne></tef:presidentJury><tef:membreJury><tef:nom>Sentieys</tef:nom><tef:prenom>Olivier</tef:prenom><tef:autoriteInterne>intervenant_1</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">061585920</tef:autoriteExterne></tef:membreJury><tef:membreJury><tef:nom>Pouli</tef:nom><tef:prenom>Tania</tef:prenom><tef:autoriteInterne>intervenant_3</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">181533065</tef:autoriteExterne></tef:membreJury><tef:membreJury><tef:nom>Termier</tef:nom><tef:prenom>Alexandre</tef:prenom><tef:autoriteInterne>intervenant_4</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">13741689X</tef:autoriteExterne></tef:membreJury><tef:membreJury><tef:nom>Filip</tef:nom><tef:prenom>Silviu-Ioan</tef:prenom><tef:autoriteInterne>intervenant_7</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">197081525</tef:autoriteExterne></tef:membreJury><tef:membreJury><tef:nom>Popovici</tef:nom><tef:prenom>Emanuel</tef:prenom><tef:autoriteInterne>intervenant_8</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">28593726X</tef:autoriteExterne></tef:membreJury><tef:rapporteur><tef:nom>Novo</tef:nom><tef:prenom>David</tef:prenom><tef:autoriteInterne>intervenant_5</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">244276455</tef:autoriteExterne></tef:rapporteur><tef:rapporteur><tef:nom>Dinechin</tef:nom><tef:prenom>Florent de</tef:prenom><tef:autoriteInterne>intervenant_6</tef:autoriteInterne><tef:autoriteExterne autoriteSource="Sudoc">060154012</tef:autoriteExterne></tef:rapporteur>
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
                        
                        <tef:ecoleDoctorale>
       <tef:nom>MATISSE</tef:nom><tef:autoriteInterne>ecoleDoctorale_1</tef:autoriteInterne>
       
       <tef:autoriteExterne autoriteSource="Sudoc">267602553</tef:autoriteExterne>
      </tef:ecoleDoctorale>
                        <tef:partenaireRecherche type="laboratoire">
       <tef:nom>
INRIA-RENNES
</tef:nom><tef:autoriteInterne>partenaireRecherche_1</tef:autoriteInterne>
       
       <tef:autoriteExterne autoriteSource="Sudoc">
133175863
</tef:autoriteExterne>
      </tef:partenaireRecherche>
                        <tef:oaiSetSpec>ddc:004</tef:oaiSetSpec>
                        
                        
                        
                    <tef:MADSAuthority authorityID="intervenant_1" type="personal"><tef:personMADS><mads:namePart type="family">Sentieys</mads:namePart><mads:namePart type="given">Olivier</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_2" type="personal"><tef:personMADS><mads:namePart type="family">Ménard</mads:namePart><mads:namePart type="given">Daniel</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_3" type="personal"><tef:personMADS><mads:namePart type="family">Pouli</mads:namePart><mads:namePart type="given">Tania</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_4" type="personal"><tef:personMADS><mads:namePart type="family">Termier</mads:namePart><mads:namePart type="given">Alexandre</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_5" type="personal"><tef:personMADS><mads:namePart type="family">Novo</mads:namePart><mads:namePart type="given">David</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_6" type="personal"><tef:personMADS><mads:namePart type="family">Dinechin</mads:namePart><mads:namePart type="given">Florent de</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_7" type="personal"><tef:personMADS><mads:namePart type="family">Filip</mads:namePart><mads:namePart type="given">Silviu-Ioan</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="intervenant_8" type="personal"><tef:personMADS><mads:namePart type="family">Popovici</mads:namePart><mads:namePart type="given">Emanuel</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="thesis.degree.grantor_1" type="corporate"><tef:personMADS><mads:namePart>Université de Rennes</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="ecoleDoctorale_1" type="corporate"><tef:personMADS><mads:namePart>MATISSE</mads:namePart></tef:personMADS></tef:MADSAuthority><tef:MADSAuthority authorityID="partenaireRecherche_1" type="corporate"><tef:personMADS><mads:namePart>
INRIA-RENNES
</mads:namePart></tef:personMADS></tef:MADSAuthority></tef:thesisAdmin>
                </mets:xmlData>
            </mets:mdWrap>
        </mets:techMD><mets:techMD ID="file_1"><mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_tech_fichier"><mets:xmlData><tef:meta_fichier>
     <tef:encodage>ASCII</tef:encodage>
     <tef:formatFichier>PDF</tef:formatFichier>
     
     
     
     <tef:taille>14307557</tef:taille>
    </tef:meta_fichier></mets:xmlData></mets:mdWrap></mets:techMD>
        
        <mets:rightsMD ID="dr_expr_thesard">
            <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_droits_auteur_these">
                <mets:xmlData>
                    <metsRights:RightsDeclarationMD>
                        <metsRights:Context CONTEXTCLASS="GENERAL PUBLIC">
                            <metsRights:Permissions DISCOVER="true" DISPLAY="true" COPY="true" DUPLICATE="true" MODIFY="false" DELETE="false" PRINT="true"/>
                        </metsRights:Context>
                    </metsRights:RightsDeclarationMD>
                </mets:xmlData>
            </mets:mdWrap>
        </mets:rightsMD>
        <mets:rightsMD ID="dr_expr_univ">
            <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_droits_etablissement_these">
                <mets:xmlData>
                    <metsRights:RightsDeclarationMD>
                        <metsRights:Context CONTEXTCLASS="GENERAL PUBLIC">
                            <metsRights:Permissions DISCOVER="true" DISPLAY="true" COPY="true" DUPLICATE="true" MODIFY="false" DELETE="false" PRINT="true"/>
                        </metsRights:Context>
                    </metsRights:RightsDeclarationMD>
                </mets:xmlData>
            </mets:mdWrap>
        </mets:rightsMD>
        <mets:rightsMD ID="dr_version">
            <mets:mdWrap MDTYPE="OTHER" OTHERMDTYPE="tef_droits_version">
                <mets:xmlData>
                    <metsRights:RightsDeclarationMD>
                        <metsRights:Context CONTEXTCLASS="GENERAL PUBLIC">
                            <metsRights:Permissions DISCOVER="true" DISPLAY="true" COPY="true" DUPLICATE="true" MODIFY="false" DELETE="false" PRINT="true"/>
                        </metsRights:Context>
                    </metsRights:RightsDeclarationMD>
                </mets:xmlData>
            </mets:mdWrap>
        </mets:rightsMD>
    </mets:amdSec>
    <mets:fileSec>
  <mets:fileGrp ID="FGrID1" USE="archive"><mets:file ID="FID1" ADMID="file_1" MIMETYPE="application/pdf" USE="maitre"><mets:FLocat LOCTYPE="URL" xlink:href="https://ged.univ-rennes1.fr/nuxeo/site/esupversions/7810be56-fee4-42fb-b300-32258d8ef1c6"/></mets:file></mets:fileGrp>
 </mets:fileSec>
    <mets:structMap TYPE="logical">
        <mets:div DMDID="desc_expr" ADMID="dr_expr_thesard dr_expr_univ admin_expr" TYPE="THESE" CONTENTIDS="http://ori-oai-search.univ-rennes1.fr/uid/rennes1-ori-wf-1-20080/oeuvre">
            <mets:div ADMID="dr_version" TYPE="VERSION_COMPLETE" CONTENTIDS="http://ori-oai-search.univ-rennes1.fr/uid/rennes1-ori-wf-1-20080/oeuvre/version">
                <mets:div DMDID="desc_edition" TYPE="EDITION" CONTENTIDS="http://ori-oai-search.univ-rennes1.fr/uid/rennes1-ori-wf-1-20080/oeuvre/version/edition">
                    <mets:fptr FILEID="FGrID1"/>
                </mets:div>
            </mets:div>
        </mets:div>
    </mets:structMap>
</mets:mets>