Présentation Powerpoint

« Meetup Data Science » Mercredi 4 mars 2020

Comment l’IA démultiplie les fonctions cognitives ?

A variant of machine learning engineer is called Deep Learning engineer. This role requires deep learning knowledge in addition to the skills profile Jean-Marie PRIGENT (Modeling, Deployment, Data Engineering). It focuses on applications, usually powered by deep ML Engineer learning, such as speech recognition, natural language processing, and computer vision. Hence, it Altran Brest requires skills specific to deep learning projects such as understanding and using various neural network architectures such as fully connected networks, CNNs, and RNNs. P as Passionately curious...

Big Data and DL but not only… Drones, FPV, CV, Maker, 3D Printer, Electronics, DonkeyCar, Edge, ... linkedin.com/in/jmprigent Machine Learning Ecosystem

Machine Learning Languages: Data Processing: - Python / R / (C++) - BIG Data framework (Cloudera/HDP/Oozie/Pig/Spark/ General Machine Learning Frameworks Scala ) - Numpy - Apache Airflow, NIFI - Scikit-Learn - NLTK, Spacy Hardware Training: - CPU, GPU, TPU, Cloud Data Analysis and Visualisation tools - Distrubuted (Spark, Kubeflow) - Pandas - Federated Learning (WO - Matplotlib centralized server) - Jupyter Notebook Inference ML frameworks for neural networks - Desktop, server … modelling - Mobile - Tensorflow / Tensorboard - Edge device (VPU: Intel NCS2, - Keras GPU: Jetson Nano & TPU: Coral - Pytorch dev board, Coral stick) - (Caffe2, mxnet) What is Deep Learning ?

Deep learning is part of a broader family of machine learning methods based on artificial neural networks with representation learning. Learning can be supervised, semi-supervised or unsupervised.

Deep learning architectures such as deep neural networks, deep belief networks, recurrent neural networks and convolutional neural networks have been applied to fields including computer vision, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, bioinformatics, drug design, medical image analysis, material inspection and board game programs, where they have produced results comparable to and in some cases surpassing human expert performance

Deep Learning : “a technique for implementing Machine Learning”

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Le Deep Learning dans l’Univers de l’IA...

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Some dates in the field of IA

1980 – Kunihiko Fukushima built the ‘neocognitron’, the precursor of modern Convolutional Neural Networks. 2001 – Two researchers at MIT introduced the first face detection framework (Viola-Jones) that works in real-time. 2009 – Google started testing robot cars on roads. 2010 – Google released Goggles, an image recognition app for searches based on pictures taken by mobile devices. 2010 – To help tag photos, Facebook began using facial recognition. 2011 – Facial recognition was used to help confirm the identity of Osama bin Laden after he is killed in a US raid. 2012 – Google Brain’s neural network recognized pictures of cats using a deep learning algorithm. 2015 – Google launched open-source Machine learning-system TensorFlow. 2016 – Google DeepMind’s AlphaGo algorithm beat the world Go champion. 2017 – Apple released the iPhone X in 2017, advertising face recognition as one of its primary new features. 2018 – Alibaba’s AI model scored better than humans in a Stanford University reading and comprehension test. 2018 – Amazon sold its real time face recognition system Rekognition to police departments. 2019 – The Indian government announced a facial recognition plan allowing police officers to search images through mobile app. 2019 – The US added four of China’s leading AI start-ups to a trade blacklist. 2019 – The UK High Court ruled that the use of automatic facial recognition technology to search for people in crowds is lawful. 2025 – By this time, regulation in FR will significantly diverge between China and US/Europe. 2030 – At least 60% of countries globally will be using AI surveillance technology (it is currently 43% according to CEIP).

This is an edited extract from the Computer Vision – Thematic Research report produced by GlobalData Thematic Research. https://www.verdict.co.uk/computer-vision-timeline/

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Apprentissage supervisé

“L’objectif de l’apprentissage supervisé est d’apprendre une fonction qui, à partir d’un échantillon de données et des résultats souhaités, se rapproche le mieux de la relation entre entrée et sortie observable dans les données.” -> Y = f (X)

L’apprentissage supervisé est généralement effectué dans le contexte de la classification et de la régression.

Classification: Un problème de classification survient lorsque la variable de sortie est une catégorie, telle que «rouge», «bleu» ou «maladie» et «pas de maladie».

Régression: Un problème de régression se pose lorsque la variable de sortie est une valeur réelle, telle que «dollars» ou «poids».

source: https://le-datascientist.fr/apprentissage-supervise-vs-non-supervise

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Apprentissage non supervisé

“L’apprentissage non supervisé (Unsupervised Learning) consiste à ne disposer que de données d’entrée (X) et pas de variables de sortie correspondantes. L’objectif est de modéliser la structure ou la distribution sous-jacente dans les données afin d’en apprendre davantage sur les données.

On l’appelle apprentissage non supervisé car, contrairement à l’apprentissage supervisé, il n’y a pas de réponse correcte ni d’enseignant. Les algorithmes sont laissés à leurs propres mécanismes pour découvrir et présenter la structure intéressante des données.”

Regroupement ou clustering: l’objectif est de séparer les groupes ayant des traits similaires et de les assigner en grappes.

Association: consiste à découvrir des relations intéressantes entre des variables dans de grandes bases de données. Par exemple, les personnes qui achètent une nouvelle maison ont aussi tendance à acheter de nouveaux meubles

source: https://le-datascientist.fr/apprentissage-supervise-vs-non-supervise

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Apprentissage semi supervisé

“Les problèmes pour lesquels vous avez une grande quantité de données d’entrée (X) et que seules certaines données sont étiquetées (Y) sont appelés problèmes d’apprentissage semi-supervisés. Par conséquent, ces problèmes se situent entre l’apprentissage supervisé et l’apprentissage non supervisé”

Le Deep Learning rentre dans la catégorie “supervisé” pour la majorité des cas et plus récemment semi supervisé avec les GAN.

Le (Deep) Reinforcement Learning rentre dans la categorie non supervisé

source: https://le-datascientist.fr/apprentissage-supervise-vs-non-supervise

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent A Visual and Interactive Guide to the Basics of Neural Networks...in a nutshell

source: jay Alammar Blog

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent A Visual and Interactive Guide to the Basics of Neural Networks...in a nutshell

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent What convolution neural network see...

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Some vocab in Computer Vision tasks

4 differents tasks in CV:

- image classification - object detection - semantic segmentation - instance segmentation

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Imaging Applications

Deep Learning has applications in all sectors of activity. It is a major issue for the industrial and scientific sectors and the safety of goods and people.

Among these uses are in particular :

- Image recognition (classification, localization and segmentation), - Description of scenes, - Facial recognition (security), - Optical Character Recognition (OCR), - Content-Based Images Retrieval (CBIR) - Medical Imaging (biology, histology, radiology), - Synthetic image generation (GAN), - Emotion detection, - Detection of age and gender, …

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent NLP-NLU Applications

NLP iis becoming more democratic, previously reserved for researchers.

The advent of personal assistants (Siri, Alexa, Google Home) and its high level of adoption proves the maturity of this technology. Studies prove that using a text transcriber is 3 times faster than writing the text.

Among the cases of use are the following: - Chatbots - Spam detection / Spam filter avoidance - Sentiment analysis (consumer opinions, customer opinions) - E-reputation - Automated translation - Subtitling of video sequences (Speech-to-Text) - Voice User Interface (VUI) (Siri, Alexa, Ok Google) …

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Get datas … and Training Neural Network

tensorboard UI

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Training sample on Fashion-MNIST under Colab

Fashion Mnist

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent 3/ Inference :

a/ Desktop | server b/ Mobile c/ Edge device (low power 2-15w) d/ ARMs (only some basic features like sound triggering)

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent … But the REAL ML Pipeline is ...

- Collect and prepare data - Developing a model - Training an ML model on the data : “Only 5-15% of Training the model the ML pipeline Evaluate the accuracy of the model Setting the hyperparameters is ML/DL“ ! - Deploy the driven model Andrew Ng - Send prediction queries to the model : Online prediction Batch Prediction - Monitor predictions continuously - Manage models and versions

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent A as ART

1/ Autodraw Autodraw: find object shape

Autodraw

2/ Style Transfer

Artiste Fred: Toile

Les toiles de Fred

file:///Users/jmp/Desktop/prez_bia/1_style_transfer/Core1_Introduc tions/neural_style_fred.html

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent RT as Real time counting

YOLO V3

It took 1.958 seconds to detect the objects in the image.

Number of Objects Detected: 20

Objects Found and Confidence Level:

1. person: 1.000000 2. person: 1.000000 3. person: 1.000000 4. person: 1.000000 5. person: 1.000000 6. person: 1.000000 7. person: 1.000000 8. person: 1.000000 9. person: 1.000000 10. person: 1.000000 11. person: 1.000000 12. boat: 0.998270 13. person: 0.999997 14. person: 0.999981 15. person: 0.999952 16. person: 0.999981 17. person: 0.999994 18. person: 1.000000 19. person: 1.000000 20. person: 0.999979

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Detection en temps reel

YOLO V3 by Joseph Redmon

> Demo mobile Yolo

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Teachable machine by Google https://experiments.withgoogle.com/teachable-machine

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent P as Pose Estimation

Demo pose estimation inference directly in browser

35FPS with tensorflow js

Pose estimation with TensorFlow.js

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent S as Sons: create sound with image only

Imaginary soundscape by Google: http://www.imaginarysoundscape.net/#/upload

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as Driver Security (outside)

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent M as Mobility

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent A as Attention : Driver Security (inside)

Drowiness and Distraction detector

source: https://github.com/incluit/OpenVino-Driver-Behaviour

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent PT as Pedestrian tracking

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent PT as Pedestrian counter

Single Shot Detector (SSD) with Opencv DNN

source: https://www.pyimagesearch.com/2018/08/13/opencv-people-counter/

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent T comme Tracking

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Image captioning: CV and NLP...

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent H as high accuracy and celerity

Detectron2 by Facebook

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as Driver assist

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent S as Security: weapons detection

Collection of handgun images, procured and RetinaNet, a deep learning-based object published by Olmos et al. in their 2018 detection architecture that seeks to publication, “Automatic handgun detection combine both the speed of one-stage alarm in videos using deep learning” detectors (ex., YOLO and SSD) with the accuracy of slower two-stage detectors (ex., Faster R-CNN)

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent CD as Cancer Identification

Mask R-cnn Train a Mask R-CNN instance segmentation network to automatically detect skin lesions, a first step in cancer identification. The ISIC Skin Lesion Dataset

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent P Pills classification

Train a Mask R-CNN instance segmentation network to automatically detect pills.

Mistakes surrounding prescription medication can and do happen, resulting in billions of euros in insurance claims, hospital bills, and of course, the incalculable value of a human.

In France, with D.I.N, clinics and hospital needs CV assisted Robot with The ISIC Skin Lesion Dataset DL features -> Brest CHU

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent F as Fish Classifier

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent F as Fraud Detection

Training set New Data

Training Inference

source: Manning MEAP: “Deep Learning with Structured Data”

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent S comme Summarization

- Bert language model - Bert Extractiver summrization - Neural coreference : coreference is the fact that two or more expressions in a text – like pronouns or nouns – link to the same person or thing

Capgemini et Altran Technologies s'unissent. Les deux géants ont conclu un accord de négociations exclusives en vue de l’acquisition par Capgemini d’Altran dans le cadre d’une OPA amicale à 14 euros par action Altran, payables en numéraire. Le montant total de la transaction s’élèvera à 3,6 milliards d’euros, avant prise en compte de la dette financière nette d’environ 1,4 milliard d’euros. "Ce rapprochement aura un impact immédiatement relutif, évalué à plus de 15% sur le résultat normalisé par action avant mise en œuvre des synergies ", expliquent les deux groupes, qui ajoutent qu'en "2023, après prise en compte des synergies, la relution devrait dépasser 25%".

L'accord a été approuvé à l'unanimité par les conseils d'administration de Capgemini et d'Altran. Par ailleurs, Capgemini a d'ores et déjà signé un accord définitif pour l'acquisition d'un bloc de 11% du capital d'Altran auprès d'actionnaires autour d'Apax Partners. L'opération vise la création d'un groupe de 17 milliards d'euros de chiffre d'affaires et de plus de 250 000 collaborateurs par le rapprochement d'un spécialiste du conseil et des services informatiques et d'un spécialiste des services d'ingénierie et de R&D. La finalisation de ce rapprochement est envisagée d'ici la fin de l'exercice 2019.

Capgemini indique avoir sécurisé un financement relais de 5,4 milliards d'euros, comprenant l'achat des titres et le rachat de la dette brute de 1,8 milliard d'euros. "Ce rapprochement, cette OPA amicale, n'est pas un changement de stratégie pour nous. On l'a appuyé avec beaucoup d'enthousiasme et de passion. Cela crée un superchampion", s'est réjoui le PDG d'Altran Dominique Cerutti lors d'une conférence téléphonique, rapporte l'AFP. "L'expertise de Capgemini et celle d'Altran sont parfaitement complémentaires et proposeront une combinaison unique sur le marché. Les entreprises sont très friandes d'innovation et nous allons accélérer grâce à notre nouvelle taille critique", a renchéri le patron de Capgemini Paul Hermelin.

Cette OPA entre le champion des entreprises de services numériques et celui du conseil en technologies devrait bouleverser le paysage de la "tech" française. Paul Hermelin a exclu que Capgemini ait besoin de céder des actifs pour des raisons de concurrence après cet accord. "On n'est pas sur des parts de marchés qui provoquent ce genre de demandes", a-t-il indiqué, rapporte l'AFP.

Grâce à l'opération, les deux groupes visent des synergies de coûts et de modèles opérationnels estimés "entre 70 et 100 millions d'euros avant impôts en année pleine" et un chiffre d'affaires annuel additionnel compris entre "200 et 350 millions d'euros" d'ici trois ans.

{ "summary": "Capgemini et Altran Technologies s'unissent. L'opération vise la création d'un groupe de 17 milliards d'euros de chiffre d'affaires et de plus de 250 000 collaborateurs par le rapprochement d'un spécialiste du conseil et des services informatiques et d'un spécialiste des services d'ingénierie et de R&D. La finalisation de ce rapprochement est envisagée d'ici la fin de l'exercice 2019. Grâce à l'opération, les deux groupes visent des synergies de coûts et de modèles opérationnels estimés \"entre 70 et 100 millions d'euros avant impôts en année pleine\" et un chiffre d'affaires annuel additionnel compris entre \"200 et 350 millions d'euros\" d'ici trois ans." }

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Launched by Ifremer in March 2017, the citizen science project “Espion des Grands Fonds”

Help our local biologists by playing deepseaspy! Need HELP www.deepseaspy.com

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Launched by Ifremer in March 2017, the citizen science project “Espion des Grands Fonds”

Help our local biologists by playing deepseaspy. www.deepseaspy.com

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Mussels Annotations

Ifremer's Deep Environment Laboratory studies the evolution of mussel beds using submerged cameras that take one photo per hour over ten years to measure their evolution.

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Mussels Segmentation

inference: automatic detection of Mussels with Mask R-cnn

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as DeepFake

“Deepfakes are fake videos or audio recordings that look and sound just like the real thing”

-> beware of the deepfake !!!

live deepfake generation with webcam feed

extract facial landmarks

source video

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as Dance

“Everybody Dance Now” (2019-08-29)

- Pose estimation - Pose normalisation - Mapping normalized pose stick figures to Target subject

source: https://arxiv.org/pdf/1808.07371.pdf

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as DonkeyCar

➔ DonkeyCar is a high level self driving library written in Python. It was developed with a focus on enabling fast experimentation and easy contribution. It is has been called the Hello World of autonomous vehicles.

➔ It is open source with an active community and support channel.

➔ Active users around the world include tinkerers, educators, educational institutions and companies.

➔ Main site: www.donkeycar.com

★ Open Source, Python ➔ Github: https://github.com/autorope/donkeycar

★ OpenCV, TensorFlow, Keras, CNN, Deep Learning ➔ Community: Discord or Slack ★ Raspberry Pi, Nvidia Jetson, GPU, Google Coral TPU ★ Behavioral Cloning, Data Augmentation ★ Transfer Learning, Reinforcement Learning

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent D as DonkeyCar

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Q as Questions ....

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Extra Content

Machine learning for creators but not only

https://runwayml.com/

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent source: source: artificial-intelligence/glossary

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent P as Papers … a lot ...

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent M comme Mobility https://selfdrivingcars.mit.edu/deeptraffic/

Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent Meetup Data Sciences - Brest is IA - 2020-03-04 - Jean-Marie Prigent