1

2017

Deep learning for aerospace applications

Alexandre Boulch

2

2017

Deep Learning

Lee Sedol 2015/10Ke Jie 2017/05

3

2017

Deep Learning

Personal assistantPersonalised learningRecommendationsRéponse automatique

Deep learning and Big data for

cardiology

4

2017

Deep Learning

5

2017

Overview

Machine Learning

Deep Learning

DeLTA

6

2017

AI

The science and engineering of making intelligent machines.

Logical, search, pattern recognition, planning, inference, learning from experience...

7

2017

AIIntelligent machines

Machine LearningLearning from experience

Model withparametersInput Output

Tuning parameters

Trainingdata

8

2017

Machine learning starts in the 60’s

9

2017

Overview

Machine Learning

Deep Learning

DeLTA

10

2017

Deep learning

11

2017

Deep learning

12

2017

Deep learning

DatasetsCompetitionEmulation

New optimizersNew neural layersNew architectures

...

BetterFaster Stronger

13

2017

Deep learning

2017

14

2017

Deep learning

AIIntelligent machines

Machine LearningLearning from experience

Deep learningAuto-learningDeep neural networks

15

2017

Deep learning

Data

Feature extraction Decision

Expert

Machine learning

Knowledge about data and application

Knowledge about statistics

16

2017

Deep learning: a massively data driven approach

Data

Features and decision

Data knowledge Statistics, optimization

Deep neural networks

Network suited for applications

17

2017

Deep learning: a massively data driven approach

Data

Features and decision

Data knowledge Statistics, optimization

Deep neural networks

Network suited for applications

18

2017

Machine learning at ONERA

DTIS

DAAA

DMAS

DMPE

DEMR DOTA

DPhiEE

19

2017

Deep learning at ONERA

DTIS

DAAA

DMAS

DMPE

DEMR DOTA

DPhiEE

A. Chan-Hon-Tong, S. Herbin, B. Le Saux, A. Boulch ...

20

2017

Semantic Map labeling

+

RG

BC

ompo

site

3 x

( 3x

3 c

onv.

+ R

eLU

)

Aerial images, multimodal (RGB, IR, DSM, ...)Fusion networksPhD Nicolas Audebert (nicolas.audebert.at)

21

2017

Point cloud labeling

2 x ( 3x3 conv. + ReLU )

Max Pooling 2x2

Max Pooling 2x2

Max Pooling 2x2

Max Pooling 2x2

2 x ( 3x3 conv. + ReLU )

3 x ( 3x3 conv. + ReLU )

3 x ( 3x3 conv. + ReLU )

Deconv. 3x3

Batch Norm2 x ( 3x3 conv. + ReLU )

Batch Norm3 x ( 3x3 conv. + ReLU )

Batch Norm3x(3x3 conv. + ReLU )

Batch Norm2 x ( 3x3 conv. + ReLU )

VGG 16

Deconv. 3x3

Deconv. 3x3

Deconv. 3x3

3 x ( 3x3 conv. + ReLU )

Conca

tenati

on

Leader on Semantic 8 LIDAR datasetTransfer to photogrammetryCode available online (DeLTA website)

22

2017

RGB and Depthfor persondetectionimprovment

PhD Joris Gueryjorisguerry.fr

Detection

23

2017

Detection in low resolutionImagesExploitation of imagesSequences for detection

Juliette Chataigner (Intern)

Detection

24

2017

Depth from defocus

Sensor specific processing Depth from de focus.PhD Macella Carvalho

25

2017

Zero Shot Learning

Zero Shot LearningLearning based on attributesPhD Maxime Bucher

Zero-Shot Learning via Visual AbstractionStanislaw Antol, Larry Zitnick, Devi Parikh

26

2017

Overview

Machine Learning

Deep Learning

DeLTA

27

2017

Deep learning for aerospace ONERA⇒

28

2017

Data

Frameworks

Development

R&D for aerospace and defense

Deep learning for aerospace ONERA⇒

29

2017

LE PRF DeLTA

Deep Lab

Applications

Core skills

30

2017

Domain adaptation

New architectures

State of the art

Databases

Core skills

31

2017

LE PRF DeLTA

Software and help to solve problems with

machine learning

Tutorials

Code

Networks

Results

Datasets

Deep Lab

Academics and industrials

Improve ONERA research exposure

32

2017

Deep Lab

Generic code

Problem code

Network base

Networks Database code

Database

Validation

How To

Learning

Practical case

33

2017

DAAAFluid

mecanics

DTISRobotics

DMASMaterial

mecanics

DEMRDetection and

recognition

DOTAAtmospheric

Corr.

Deep Lab

Applications

Core skills

34

2017

TransferExperimentEvaluate

Move forwardInnovate

CapitalizeShareReproduce

Deep Lab

Applications

Core skills

4 year project

delta-onera.github.io

35

2017

“We chose it because we deal with huge amounts of data. Besides, it sounds really cool.”

Larry Page - Google