L intelligence artificielle et l apprentissage profond appliqués à la maintenance prédictive

Save this PDF as:
 WORD  PNG  TXT  JPG

Size: px
Start display at page:

Download "L intelligence artificielle et l apprentissage profond appliqués à la maintenance prédictive"

Transcription

1 L intelligence artificielle et l apprentissage profond appliqués à la maintenance prédictive Stéphane Canu, INSA Rouen Normandy University asi.insa-rouen.fr/enseignants/~scanu Tech Hour Normandie AeroEspace 19 décembre 2017

2 Road map 1 L intelligence artificielle et l apprentissage automatique 2 L apprentissage profond (Deep learning) Du neurone aux réseaux profonds ImageNet La mode du deep 3 Quoi de neuf avec le deep learning? Big is beautiful Apprentissage et optimisation Architectures profondes 4 Conclusion

3 Airplane Health Management Safety inspection inspection visuelle tap test : Contrôle santé des structures composites Inspecter grace à ses sens voir, écouter et sentir (toucher) Thèse de Meriem GHRIB

4 Automated visual inspection Automated image processing aircobot.akka.eu/, donecle.com/about-donecle/ Automated exterior inspection of an aircraft with a pan-tilt-zoom camera mounted on a mobile robot, 2015 Apprendre à voir, écouter et sentir

5 Human vs. machine learning Child s learning capacities to walk: one year to speak: two years to think: the rest of my life

6 Apprendre à penser

7 Deux types d intelligence artificielle (AI) Artificial general intelligence (AGI), long term AI, short term AI, applied AI, Deep learning

8 Machine learning definition Machine Learning (T. Mitchell, 2006) A computer program CP learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E Key points experience E : data performance measure P : optimization tasks T : utility automatic translation play chess or go...do what humans do

9 Exemples of image processing related learning tasks T Recognize Predict Detect Read Localize Scene analysis

10 Exemples of image processing related learning tasks T Recognize Predict Detect Read Localize Scene analysis A single device can solve them all: our brain

11 Learn the features together with the prediction Zhao, Rui, et al. "Deep Learning and Its Applications to Machine Health Monitoring: A Survey." arxiv preprint (2016).

12 So far so good Machine learning is Example-based programming task data performance A lot of applications in computer vision (as human learn to see) detection recognition localization captioning... Deep learning the dream of using a single multi task device for all vision tasks feature learning is a challenge addressed by deep learning on many application deep learning give the best results

13 Road map 1 L intelligence artificielle et l apprentissage automatique 2 L apprentissage profond (Deep learning) Du neurone aux réseaux profonds ImageNet La mode du deep 3 Quoi de neuf avec le deep learning? Big is beautiful Apprentissage et optimisation Architectures profondes 4 Conclusion

14 Le Neurone biologique

15 Le Neurone formel (McCulloch & Pitts, 1943) Défini un hyperplan x 2 w t x + b = 0 x 1, w 1 x 2, w 2... x p, w p 1, b h y = ϕ(w t x + b) x 1 x entrée IR p w poids synaptiques, b biais ϕ fonction d activation y sortie IR ϕ(t) = tanh(t)

16 Le perceptron (Rosenblatt, 1957) p f (x) = H ϕ j (x)w j + w 0 j=1 Règle du Perceptron Tant que on n a pas convergé : 1 tirer un exemple (x i, y i ) 2 calculer la prédiction f (x i ) 3 adaptater les poids w w + ρ 2 (y i f (x i ))ϕ(x i ) Du point de vue optimisation n ˆR n (w) = max(0, y i w ϕ(x i )) i=1 Méthode de gradient stochastique

17 Non linearity combining linear neurons: the Xor case Alpaydın, Introduction to Machine Learning, 2010

18 Neural networks as universal approximator Running several neurons at the same time y = ϕ(w 3 h (2) ) h (2) = ϕ(w 2 h (1) ) h (1) = ϕ(w 1 x) x Multilayered neural networks in layers Use backpropagation to learn internal representation W 1, W 2, W 3 from L. Arnold PhD

19 The Asimov Institute:

20 OCR: the MNIST database (Y. LeCun, 1989) use convolution layers

21 The caltech 101 database (2004) 101 classes, 30 training images per category...and the winner is NOT a deep network dataset is too small use convolution + Recitification + Normalization + Pooling in What is the Best Multi-Stage Architecture for Object Recognition? Jarrett et al, 2009

22 Deep networks, big computers: use GPU (2009) Large-scale deep unsupervised learning using graphics processors R Raina, A Madhavan, AY Ng - Proceedings of the ACM, 2009

23 Big Data : the image net database (Deng et al., 2012) ImageNet = 15 million labeled high-resolution images of 22,000 categories. Large-Scale Visual Recognition Challenge (a subset of ImageNet) 1000 categories. 1.2 million training images, 50,000 validation images, 150,000 testing images.

24 A new fashion in image processing shallow approaches deep learning Y. LeCun StatLearn tutorial

25 karpathy s blog: karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/ ImageNet results 35 top-5 error in % 152 layers number of layers 30 human performance publication year 2012 Alex Net 2013 ZFNet 2014 VGG 2015 GoogLeNet / Inception 2016 Residual Network

26 Deep architecture and the image net (15%) The Alex Net architecture [Krizhevsky, Sutskever, Hinton, 2012] Convolution + Recitification (ReLU) + Normalization + Pooling 60 million parameters using 2 GPU 6 days regularization data augmentation dropout weight decay

27 From 15% to 7%: Inceptionism Network in a network (deep learning lecture at Udacity) Christian Szegedy et. al. Going deeper with convolutions. CVPR 2015.

28 From 7% to 3%: Residual Nets. Beating the gradient vanishing effect K. He et al, 2016

29 Learning Deep architecture min W IR d n f (x i, W ) y i 2 + λ w 2 i=1 d = n = 1, λ = f is a deep NN Y. Bengio tutorial

30 Deep learning and the industry M. Zukerberg show up at a deep learning research workshop (2013) the GAFAM they got the infrastructure (hard+software) they got the data deep learning bridges the gap between applications and ML Deep Learning As A Service a lot of available API (google, microsoft, Nvidia, Amazon... )

31 Deap learning applications image audio speech text (NLP) translation robotics playing games science (Higgs Boson) Mostly related with specialized low level perception

32 So far so good from the formal neuron to deep learning one neuron is a linear perceptron many layered neurons are non linear multilayered perceptrons deep networks is a new name for multilayered perceptrons deep learning breakthrough starts with ImageNet better than human performances on many perception tasks deep learning could transform almost any industry the AI revolution Neural networks+backpropagation exist since 1985 what s new?

33 Road map 1 L intelligence artificielle et l apprentissage automatique 2 L apprentissage profond (Deep learning) Du neurone aux réseaux profonds ImageNet La mode du deep 3 Quoi de neuf avec le deep learning? Big is beautiful Apprentissage et optimisation Architectures profondes 4 Conclusion

34 What s new with deep learning a lot of data (big data) big computing resources (hardware & software), big model (deep vs. shalow) new architectures new learning tricks from Recent advances in convolutional neural networks Gu et al. Pattern Recognition, 2017

35 Dealing with a lot of data. ImageNet: 1,200,000x256x256x3 (about 200GB) block of pixels Image understanding MS COCO for supervised learning YFCC100M for unsupervised learning Visual genome: data + knowledge Video understanding: Youtube 8M (1.71 Terabytes) Climate monitoring with HD images (about 15 TB and DL at 15PF)

36 Dans l industrie aéronautique crédit HP,

37 Andrew Ng basic recipe for machine learning Why is Deep Learning taking off? fuel = data engine = model (a deep network). Andrew Ng GTC 2015 Keynote, GPU Technology Nvidia

38 GPU needed Now 2 hours with Nvidia DGX-1, and enough Memory Yann LeCun: learning a relevant model takes 3 weeks

39 Deep learing frameworks Tensoflow (Google) is the most popular with Keras

40 How to start with deep learning? Andrej Karpathy, Deep Learning Summer School 2016

41 Success story: Updating Google Maps with Deep Learning Requirements Installed TensorFlow library 158Gb to download FSNS dataset: 16Gb of RAM (32Gb is recommended) training 60 h with GPU Titan X to train from scratch: python train.py to train a model using a pre-trained inception weights as initialization: wget tar xf inception_v3_2016_08_28.tar.gz python train.py --checkpoint_inception=inception_v3.ckpt

42 What s new with deep learning from Mastering the game of Go without human knowledge D. Silver et al. Nature, 550, 2017

43 Road map 1 L intelligence artificielle et l apprentissage automatique 2 L apprentissage profond (Deep learning) Du neurone aux réseaux profonds ImageNet La mode du deep 3 Quoi de neuf avec le deep learning? Big is beautiful Apprentissage et optimisation Architectures profondes 4 Conclusion

44 Conclusions so far so good deep learning major breakthrough better than human on low level perception tasks data needed complex algorithms the future of deep learning data: the new energy hard to compete with GAFAMs unsupervised learning (Generative Adversarial Networks) transfer learning (representation learning)

45 Pendant ce temps en Normandie, au LITIS, au GREYC 200 chercheurs en Normandie Universités et écoles d ingénieurs litislab.fr, greyc.fr NormaSTIC: fédération CNRS 20 en machine learning 20 en traitement d image Applications {x i, y i } Chemoinformatics BCI (interfaces cerveaux machines) Audio détection de piétons... Deep learning Boites à outil, CRIAN Deep in France (soutenu par l ANR) optimisation et ML, transport optimal et GAN

46 Shameless advertising: june in Normandy

47 To go further books I. Goodfellow, Y. Bengio & A. Courville, Deep Learning, MIT Press book, Gitbook leonardoaraujosantos.gitbooks.io/artificial-inteligence/ conferences NIPS, ICLR, xcml, AIStats, Journals JMLR, Machine Learning, Foundations and Trends in Machine Learning, machine learning survey lectures Deep Learning: Course by Yann LeCun at Collège de France in 2016 college-de-france.fr/site/en-yann-lecun/inaugural-lecture h00.htm Convolutional Neural Networks for Visual Recognition (Stanford) deep mind ( CS 229: Machine Learning at stanford Andrew Ng Blogs Andrej Karpathy blog (