Sparse Coding
Transcription
Sparse Coding
1 DYNamiques de l'Information 2 PR, 4 MC, 1 postdoc, 2 IG, 2+2 doctorants fin 2015 11 PhD + 1 HDR soutenues sur 2010-2015 Axe : Signal → Apprentissage / Structuration → Information Recherches théoriques : Optimisation convexe régularisée, Classification semi/supervisée, Apprentissage génératif de données fonctionnelles, Réseaux profonds (convolutionels), Modèles neuromimétiques, Sparse Coding, Big-Data, Modélisation et manipulation de données Valorisation : Fossé sémantique sur grande échelle, annotation SON, image, vidéo, surveillance environnementale, RI Big Data, veille web. 2 Dyni = Interdisciplinary Research on Multimodal Signal Methods for data structuration • Graph Sparse Coding, Sparse Coding • Bayesian Non Parametric clusterin • Bayesian user model... – Image classification – Acoustic classification (nevus, web, handwritten, speech, bioacoustics) – Ontology / Data user Model Interface entre les mathématiques appliquées (statistique) et l'informatique avec un intérêt particulier pour le traitement du signal, et l'apprentissage statistique Depuis 2011 : 1) Segmentation de grandes données Temporelles ; 2) Analyse de données fonctionnelle; 3) Régularisation Bayésienne ; 4) Apprentissage Bayésien non-paramétrique; 5) Modélisation robuste de données non-Gaussiennes et atypiques ; 6) Tracking de sources concurrentes ; 7) Personnalisation RI 3 Projets principaux 2010-15 Projets ANR : ● SPID (2015-16) : Surveillance acoustique passive, application à suivi de drone (67 K) ● Cognilego CONTINT (2011-14) : Réseau profond pour la lecture (80 K) (PI) ● ANCL BLANC (2009-12) : Modèle neuromimétiques de la lecture (100 K) ● AVEIR MDCA (2008-11) : Annotation d'image du web (80 K) Autres : ● RAPID 'PHRASE' 2012-2015 en traitement image sous-marine (100 K) ● FUI 'SYCIE' 2014-2017 en traitement donnés drone sous marin (100 K) ● Ministère Environnement : Bioacoustique sur Parc National Port-Cros : VAMOS (2015-17), DECAN (2015), DECAV (2013-15) (total 140 K) - Animation nationale : 'Scaled Acoustic Biodiversity' sabiod.org, 5 UMR et coll. Internat,(180 K sur 201215), Mission Interdisciplinaire du CNRS ; Responsable de l'Action masse de données bioacoustiques du GDR MADICS 2015-17. - Animation internationale : organisation de plusieurs challenges et workshops à ICML13, 14, NIPS13, ICDM15, LIFECLEF14,15,16... - Pilotage interlabo sur UTLN de JASON “Joint Acoustic Survey” regroupant 3 UMRs sur UTLN (50 K) - Projet indus. Sermicro (éq. IBM Espagne) (70 K) - IUF (100 K) -Thèses : CIFREs : COEXEL (110 K) en veille informationnelle ; Bio-id (100 K) dès fin 2015 ; 3 MRTs Total : ~1 M euros en 5 ans Results 2011-15 • 20 ACL (IEEE TIP, Pattern Recognition, IJPRAI, Applied acoustics, JASA, MTAP, Neurocomp., Meth. Ecology & Evolution, Science China IS,...) 40 ACTI A (ICPR, ICASSP, IJCNN...), 10% inter-équipe (IM, Escodi), 10% internat. . 5 book chapters, 2 (+1) brevets (EU, USA) • . 6 Wkp Machine Learning (IEEE EADM15, ICML13, ICML14, NIPS13, IJCNN12) . BDA 2015 (150 part.), Summer schools ERMITES (10 éd. 2007 : 10*30 part.) . Coll. : Cornell univ NY, Victoria univ (Canada), Pavia univ., Heifi univ (CN) : ENS Ulm DATA Team, INRIA Zenith, CIRAD, PNPC, MNH Paris... • • • LifeClef Lab creation with INRIA Zenith at Clef. 2013-15 Reviewers in PLOS one, IEEE TSI, Applied Acoustics, JASA... Patent Licensing EU, USA / Price : Minist. Ecology 2014, 2015 2016... . Axe information UTLN : JASON 10 (Pr + MC) Escodi & UMR IM2NP . HPC Bioacoustics : Coll. ON Canada, Pavia univ., Cornell Univ. . FUI LSIS GIPSA DCNS 2016... ? . Project Eu H2020 MOBIDIC posé (Berlin, UK) ? . Feder Alpes InterReg ? 5 9 UMR, 4 CNRS institutes (IN2SI, INSB, INEE, INSU) ; Internat. Coll : New York , Cornell univ, Cibra... Industry : Cyberio, Sermicro, Nortek, Dodotronics… Parc National Port-Cros, Pelagos, réserves intégrales Italie, île de la Réunion, Vancouver... Data.ENS GIPSA CPPM LIF Zenith LAMFA SABIOD synopsis Array sensors 7 SABIOD Crowdsourcing CNRS IN2SI & INSB & INEE (SABIOD) CNRS (INSU) = 68 years Fe = 100 Hz from 2005 to 2015 0 1 recordings 1 1 2 0 1 2 Vancouver BIRD Madagascar Réunion Nvelle Calédonie HPC system for Big Data Detection Localisation of Marine Mammals Natural sounds analyses needs Scaled Learned Representation Feature learning for fast indexing & classifying Discovering new spatio temporal pattern for (un)known sources Scaling the methods 0 1 1 1 Infinite HMM for unsupervised segmentation 2 0 1 2 Units 1, 2, ,t 0 1 1 1 2 0 1 2 Megaptera Whale song IHMM decomposition THEMA 0 1 1 1 Sentence 1 p2 p3...... Sentence 2 p2 p1 p5 p2 …. p5 2 0 1 2 Chamroukhi et al. 2014 Sparse Coding computation Why Sparse Coding ? ● More discriminatve ● Beter generalizaton for new data ● Reducton of the reconstructon error ● Data compression ● Each data vector x is expressed as a c linear combinaton of a dictonary D of i i size K (only one in usual K-means) ● Formulaton : ● introduces sparsity (regularizaton constraint : some contributon are non zero) Iteratve learning of D and C untl convergence by LASSO and K-SVD algorithms ● Complexity for projecton in ~O(K n nnz), n the number of vectors to project, nnz the average non-zero coefcients ● [ B. A. Olshausen and D. J. Field. Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision Research, 37:3311–3325,1997 ] 1 4 Classification efficiency of SC vs state of the art R [ Razik Paris and Glotin, ICPRAM 2011 ] Real data :broadcastnews 2h duration At equivalentsize SC outperformsGMM Overtrainingeffect for GMM Still increasing recognition results for SC Humpback whale song analysis by sparse coding : exploring song components • Humpback songs are structured – Most decomposition algorithms use prior information • Unsupervised determination of recurrent pattern in a data flow – Usually clustering – K-means clustering drawbacks: • Each cluster may not cover all the space • Each cluster not suit the data. • This study: – Validate the “subunit” component hypothesis – Propose a method to automatically classify song species by the “subunit” components of the song * Material : Songs recorded in Madagascar, Reunion & New Caledonia, Tonga, Hawaï... 48 kHz FS, from 2008 to 2014 (no 2010 neither 2011) 13 MFCCs, 10 ms frame shift, 32 ms frame length N windows are concatenated to desired T scale (250 ms Long term WHALE SONG EVOLUTION shown by Sparse Coding Year evolution from A to B = log( P( xy, A ) / P ( xy, B) ) D = 1Byear - A = 1 year D = 4 years yY Y X X Y D = 5 years D = 6 years x Our algorithm, by unsupervised dictionary learning of a proto-lexicon of the song of the humpback whale allows long term systemic units are efficiently extracted and show the variation of the composition of the songs from one year to another => WORLD SCALE BIOPOPULATION ANALYSIS [ Glotin et al J. Acoust. Soc. Am. 133, 3311 (2013) Doh 2014 Phd Thesis ] Analysis and display at diferent tme scales to identfy acoustc structures of singing species Methods in Passive Acoustic 3D tracking Transitivity … to filter TOA (2 patents, EU USA...) licensed Time Delay of Arrival filtering by transitivity, and Dopler constraint : I T Robust to different species, chirp, voicing,... EU, USA patent, Maturation (2013-14), licencied 2015... Scaled Tracking by Sparse Coding [Glotin, Razik et al in POMA ASA 2013] B A Replace cross. corr By cosine similarity measure : Cos(A,B) = ( A .B ) / ( ||A|| . ||B||), Higher the cosine is, the more the vectors are correlated, FAST vectorial implementation / parallel processing : allcosines( h1, h2 ) = ( H1 * H2’ ) / (norm(H1’) * norm(H2)) , where Hi is the matrix of the 1024 by 10 minutes frames, * is the matrix product, norm(Hi) is the L2 norm of each frame vector of Hi. [Glotin, Razik et al in POMA ASA 2013] Scaled Sparse Time Delay of Arrival estimations applied to Voicing of Whales (Minke) on 10 sec. recordings Direct CrossCorr(s1, s2) Sparse code correlation : Cosinus(SC(s1) ,SC(s2)) with dictionary learned on s1 U s2 [ Hervé GLOTIN - Joseph RAZIK - GIRAUDET Pascale - Sébastien PARIS - Frédéric BÉNARD Sparse coding for fast minke whale tracking with Hawaiian bottom mounted hydrophones" , International Workshop on Detection, Classification, Localization & Density Estimation of Marine Mammals using Passive Acoustics, Portland, USA, supported by ONR Dpt of the Navy & Acoustical Society of America (ASA) 2011 Time Delays Of Arrival Estmatons We extract 14 TDOA over these 30 minutes, between h1,h3,h4,h6 => Coherent and regular variations Large aperture hydrophone array : real time tracking system : demo. on real data: [ Patent Glotin et al. Multiple whale tracking USA patent 2013 Glotin et al. Whale Cocktail Party, Canac Acoustics, 2008 Bénard Glotin, Neutrino whale tracking, Applied Acoustics 2011 ] Online demo at http://sabiod.org/tv RANGE [ 500 to 5000 m] prec :15m Illustraton on large scale whale monitoring : Marseille-Toulon-Nice [ 2013 Abeille Phd, Glotin - Coll G. Pavan 2014 Doh Phd, Glotin - Coll Adam ] Multimodal Platform SABIOD BOMBYX Observatory – Univ. Toulon http://glotin.univ-tln.fr/BOMBYX South Port-Cros National Park – 2012... Collaboration with UMR LSIS, IUF, PNPC, UMR MIO, PME Osean Multimodal instrumentation : Hydrophones (stereo), ADCP, Video submarine camera Objective : Inverse model by machine learning of audio, surface, and subsurface sea-states, at short and long ranges. Application transient analysis on Physeter (cachalot) recorded on BOMBYX 2014 Distance estimation using only one hydrophone[Doh et al 2014] Physeter Whale, Hydro Distance Estim. (meter) 2400m 12 lines 25 storeys / line 3 PMT / storey 40 km to shore 450 m Junction box 60 m Readout cables Join us in the Large Scale Bird Id Focus on Brazil area Most populated country in XC dataset 1st country in terms of endangered species 14027 recordings from the top-501 species minimally 15 recordings per species (max 91) minimally 10 diferent recordists per species (max 42) Metadata: type of sound (call, song, alarm, etc) day + tme + locaton + gps + elevaton recordist & personal remarks, quality rate Life Clef 2014 Large scale challenge BIRD SPECIES from Bresil using Crowdsourcing and Crowdsolving : the largest species classifcaton system # load Giga Oct