Autonomous Systems: Deep Learning

1. Overview: From AI to Deep Learning

Prof. Dr.-Ing. Nicolaj Stache
Heilbronn University of Applied Sciences

Introduction Nicolaj Stache

X Study RWTH Aachen: Electrical
Engineering + Information Technology
X Doctorate RWTH Aachen: Industrial
Image Processing
X 2010 – 2016: Continental
X Head of Artificial Intelligence Center
X Leader for Sensorics (automated driving)
X Team Lead Camera Monitoring

X 2016: Professor of Measurement and

Sensor Technology in the automotive
field

X 2017: Program director of Automotive

Systems Engineering, HHN

X 2018: Co-Founder of Center for

Machine Learning (ZML) @ HHN

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 2

Overview on the Course

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 3

Overview on the Course

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 4

Sources of Inspiration

X Goodfellow et al: Deep Learning, The X Stanford Lectures on CNNs:

MIT Press, 2016 http://cs231n.github.io/
https://www.youtube.com/playlist?l
X Trask: Grokking Deep Learning,
ist=PLC1qU-LWwrF64f4QKQT-
Manning Publications, 2017
Vg5Wr4qEE1Zxk
X Online Books:
X http://neuralnetworksanddeeplear
X Udacity courses on deep learning:
ning.com/
X https://de.udacity.com/course/
X http://www.deeplearningbook.org/
deep-learning--ud730
X https://www.udacity.com/cour
se/deep-learning-
nanodegree--nd101

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 5

Learning Objectives

X You can describe the meaning of…

X Artificial Intelligence
X Machine Learning
X Deep learning

X You can describe the differences between

supervised, unsupervised, reinforcement learning
X You know parametric and non-parametric models

X You know where and why is deep learning successful

X You can give examples of deep learning applications

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 6

Motivation Video (German)

Source: https://www.zdf.de/nachrichten/heute-plus/deep-learning-104.html, 03.03.2018

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 7

Historical Overview

Source:
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_pa
Auto.-Sys: Deep Learning rt1.html, 21.09.2017 Prof. Dr. N. Stache 8
Historical Overview 1943

Dream of “self-thinking machine”

1943 - McCulloch & Pitts

X Re-engineer the principle of brain
X Introduction of threshold logic unit

X Realization of logic gates: ‫ݔ‬ଵ

and, or, not inputs output

‫ݕ‬
X Only binary inputs, inhibitory inputs ‫ݔ‬௡ ‫ݑ‬
binary output
X No trainable weights ‫ݔ‬௡ାଵ
inhibitory ‫ݔ‬௡ା௠
inputs
Source:
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_pa
Auto.-Sys: Deep Learning rt1.html, 21.09.2017 Prof. Dr. N. Stache 9

Historical Overview 1957 - 1969

“The perceptron is the embryo of an electronic computer

that [the Navy] expects will be able to walk, talk, see,
write, reproduce itself and be conscious of its existence.”
Rosenblatt (1957)

X Weights of the inputs

and threshold of the output
are learned
X Binary classification possible
X Single layer perceptron cannot
model XOR gate
Æ AI-Winter

Source:
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_pa
rt1.html, 21.09.2017;
Auto.-Sys: Deep Learning https://upload.wikimedia.org/wikipedia/en/5/52/Mark_I_perceptron.jpeg, Prof. Dr. N. Stache 10
03.10.2018
Historical Overview 1986 - 1995

Multilayer Perceptron + Backpropagation to learn weights

X Solution to nonlinear problems
X Multilayer Perceptron == Feedforward artificial neural network

But:
X Failed to train deep structures
X High computational effort
X Alternative Approach: Support Vector Machine
Æ 2nd AI Winter
Source:
http://deeplearning.net/tutorial/mlp.html, 03.10.2018
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_pa
Auto.-Sys: Deep Learning rt1.html, 21.09.2017
Prof. Dr. N. Stache 11

Historical Overview from 2006

Era of deep learning starts

X Breakthrough in 2012 with image recognition
and language processing

Source:
https://cdn.edureka.co/blog/wp-content/uploads/2017/05/Deep-Neural-
Network-What-is-Deep-Learning-Edureka.png, 03.10.2018
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_pa
Auto.-Sys: Deep Learning rt1.html, 21.09.2017 Prof. Dr. N. Stache 12
What is AI?

Artificial Intelligence

Definition (SOURCE: MERRIAM WEBSTER)

Machine Learning
• 1: a•branch of computer
e.g. logistic science dealing with the simulation of intelligent
regression,
behavior in computers
• system acquires own knowledge by extracting patterns from
hand-defined features, i.e. it learns from data
• 2: the capability of a machine to imitate intelligent human behavior
Representation Learning
• e.g. shallow autoencoders
• System acquires own knowledge by finding feature
representations fromtest:
1950 – Turing data and extracting patterns of it
Human
Deep evaluates natural language conversations …
Learning
• e.g. multilayercannot
… if evaluator reliably tell the machine from
perceptron
• human,
Systemthe test is passed
acquires own knowledge by finding a
hierarchy of representations and extracting patterns
of it

Auto.-Sys: Deep Learning Source: https://www.merriam-webster.com/dictionary/artificial%20intelligence Prof. Dr. N. Stache 13

31.08.2017

Example of AI – Tic Tac Toe

Minimax-Tree:
X Player x wins at maximized score, Player y wins at minimized
score

X Many problems can be

formulated as a game
X If tree-size grows too large,
pruning + heuristics are used to
only partially evaluate the
search tree.
(e.g. tic tac toe already has
255,168 leaf nodes)

Auto.-Sys: Deep Learning Source: Russel, Norvig: Künstliche Intelligenz, Pearson Prof. Dr. N. Stache 14
What is Machine Learning?

Artificial Intelligence (not completely)

0
• knowledge input by hand-designed rules
• e.g. knowledge base system, chess-playing system
Machine Learning
• system acquires own knowledge by extracting patterns from
hand-defined features, i.e. it learns from data
Representation Learning
• System acquires own knowledge by finding feature
representations of data and extracting patterns of it
Deep Learning
• System acquires own knowledge by finding a
hierarchy of representations and extracting patterns
0 of it

Auto.-Sys: Deep Learning Source: Goodfellow et. al.: Deep Learning, MIT Press, 2016. Prof. Dr. N. Stache 15

Overview on Machine Learning

X Machine learning:
“A field of study that gives computers the ability to learn without being
explicitly programmed” (Arthur Samuel)
X More specific:
X Data used to build models
X Models used to make predictions on new data

Features Result
X Example data:
Hours of Hours of Exam
study sleep passed
1 8 0
Entities, 10 6 1
Observations
Instances 20 4 0
11 10 1

X Examplary task: try to predict if exam is passed, given hours of study and
hours of sleep Æ so-called supervised learning
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 16
Overview on Machine Learning

X Examplary task: try to predict result of exam given hours of study

and hours of sleep Æ so-called supervised learning

Hours of Hours of Exam

study sleep passed
1 8 0
Entities,
10 6 1
Observations
Instances 20 4 0
11 10 1
desired output

input compare

output

model
Auto.-Sys: Deep Learning Source: Trask: Grokking Deep Learning, Manning Publications, 2017. Prof. Dr. N. Stache 17

Other Examples of Supervised Learning

X Using the pixels of an image to detect the presence or absence of

a cat
X Using the movies you've liked to predict movies you may like
X Using someone's words to predict whether they are happy or sad.
X Using weather sensor data to predict the probability of rain.
X Using car engine sensors to predict the optimal tuning settings.
X Using news data to predict tomorrow's stock price.
X Using an input number to predict a number double its size.
X Using a raw audio file to predict a transcript of the audio.

Auto.-Sys: Deep Learning Source: Trask: Grokking Deep Learning, Manning Publications, 2017. Prof. Dr. N. Stache 18
Two Main Applications of
Supervised Learning passed
Hours of study
failed
X Classification
X find the class boundaries
X output is categorical
(+ confidence),
e.g. exam passed, exam failed
Hours of sleep

X Regression Head circumference

observations
X Find the underlying
relationship
X Output is a real value
X e. g. head circumference in cm

Body size
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 19

Parametric ML-Algorithms
passed
Hours of study
failed
X Parametric
X Model with fixed number of
parameters, number is
independent from data
X e.g.: model of a line
Hours of sleep
X Pro:
X Simple
X Fast to determine parameters (= train)
X Only few data is required to fit the model
X Con:
X Model is constrained to specific form
X Complexity is limited by number of parameters Æ could lead to
poor fit

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 20

Nonparametric ML-Algorithms
passed
Hours of study
failed
X Nonparametric
X No a-priori assumption about
the form of the model function ?
X No fixed amount of parameters
X Example: ݇-nearest-neighbors
Hours of sleep
algorithm
X Pro:
X Flexibility, large number of functional forms to model
X No or only weak assumptions about the underlying function
needed
X Con:
X Often: more training data required
X More risk to overfitting

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 21

Overfitting

X Goal:
Find model, using given data Æ make predictions on new data
X Example: KNN-Classification
Bildqquelle: https://kevinzakka.github.io/2016/07/13/k-nearest-neighbor/ (28.01.2018)

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 22

Overfitting

X Goal:
Find model, using given data Æ make predictions on new data
X Example: Regression

Training data
Real data
Model capacity: 2 Model capacity: 6
‫ݕ‬ ‫ݕ‬
Training error: 0.5 Training error: 0.1
Test error: 0.5 Test error: 2.3

Î “Overfitting”

‫ݔ‬ ‫ݔ‬

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 23

Underfitting

X Goal: Find model, using given data Æ make

predictions on new data
X Example: Regression

Training data
Model capacity: 3 Real data Model capacity: 2
‫ݕ‬ Training error: 0.1 ‫ݕ‬ Training error: 3.8
Test error: 0.1 Test error: 3.9

Î “Underfitting”

‫ݔ‬ ‫ݔ‬

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 24

Overfitting, Underfitting

X The more parameters of a parametric model, the higher the

model’s capacity
X Recognize Overfitting by checking the algorithm with unseen
validation data Æ evaluate generalization error

Auto.-Sys: Deep Learning Source: Goodfellow et. al.: Deep Learning, MIT Press, 2016. Prof. Dr. N. Stache 25

Summary and Overview on Learning

Supervised Unsupervised Reinforcement

Data: Data: Data:
• Input data • Input data • Input data
• Output ground truth + Reward function

Scope: Scope: Scope:

• Imitate the pattern • Find patterns in the • Tries to find out what
between input and input data e.g. to do to maximize
output Æ predict the clustering of input reward
output data

9
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 26
Unsupervised Learning

X Only input data available Features Result

Hours of Hours of Exam

study sleep passed
1 8 0
Entities,
10 6 1
Observations
Instances 20 4 0
11 10 1

X Algorithms try to find structure in the data

X Main Application:
Clustering of data (i.e. find inherent groupings in the data)
X Example: ݇-means Algorithm

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 27

Example: K-Means Algorithm

Hours of study
X Input:
X ݇ … number of clusters we
want to find
X Input data points (h_sleep, h_study)
X Approach:
X Place ݇ centroids at random locations Hours of sleep
X Repeat until convergence:
X Assign each data point to the nearest centroid Æ this forms a cluster
X Recompute centroid positions with the points of each cluster
X Stop when none of the cluster assignments change

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 28

Summary and Overview on Learning

Supervised Unsupervised Reinforcement

Data: Data: Data:
• Input data • Input data • Input data
• Output ground truth + Reward function

Scope: Scope: Scope:

• Imitate the pattern • Find patterns in the • Tries to find out what
between input and input data e.g. to do to maximize
output Æ predict the clustering of input reward
output data

9 9
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 29

Reinforcement Learning
Q-Learning
Go left: +100

X Agent performs actions in an environment

and gets points for the actions as a reward
Go right: 0

X Agent tries to learn how to best interact with the environment to

maximize reward (without having a model of the environment)
X Q-Function(state, action): predict maximum future reward
Î can be used for decision making to maximize cumulative
reward
X Q-Function can be implemented as table with state as rows,
actions as columns, containing the expected reward

Choose action Measure

Init Q Perform action
from Q reward

Update Q
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 30
Reinforcement Learning
Q-Learning

Source: https://www.youtube.com/watch?v=n7370dzCZ1o (27.01.2018)

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 31

Summary and Overview on Learning

Supervised Unsupervised Reinforcement

Data: Data: Data:
• Input data • Input data • Input data
• Output ground truth + Reward function

Scope: Scope: Scope:

• Imitate the pattern • Find patterns in the • Tries to find out what
between input and input data e.g. to do to maximize
output Æ predict the clustering of input reward
output data

9 9 9
Auto.-Sys: Deep Learning Prof. Dr. N. Stache 32
Summary and Overview on Learning

Supervised Unsupervised Reinforcement

Data: Data: Data:
• Input data • Input data • Input data
• Output ground truth + Reward function

Scope: Scope: Scope:

• Imitate the pattern • Find patterns in the • Tries to find out what
between input and input data e.g. to do to maximize
output Æ predict the clustering of input reward
output data
Classification:
predict discrete Clusters Actions
responses
Regression:
predict continuous
responses
Training BEG Prof. Dr. N. Stache 33

Deep Learning is a Section of …

Artificial Intelligence (not completely)

0
• knowledge input by hand-designed rules
• e.g. knowledge base system, chess-playing system
Machine Learning
• system acquires own knowledge by extracting patterns from
hand-defined features, i.e. it learns from data
Representation Learning
• System acquires own knowledge by finding feature
representations of data and extracting patterns of it
Deep Learning
• System acquires own knowledge by finding a
hierarchy of representations and extracting patterns
0 of it

Auto.-Sys: Deep Learning Source: Goodfellow et. al.: Deep Learning, MIT Press, 2016. Prof. Dr. N. Stache 34
Deep Learning vs. Other AI Disciplines

Legend:
Developed
Hand- by human
Input designed Output Learned by
program machine

Hand- Mapping
Input designed from Output
features features
Representation learning

Mapping
Input Features from Output
Machine Learning

features

More Mapping
Deep l.

Input Features abstract from Output

features features
AI

Auto.-Sys: Deep Learning Source: Goodfellow et. al.: Deep Learning, MIT Press, 2016. Prof. Dr. N. Stache 35

What is a Feature?

X Definition:
Individual measurable property of a phenomenon being observed
[Source: Bishop, Christopher (2006). Pattern recognition and machine learning. Berlin: Springer]

X Examples of hand-designed features in Computer Vision:

X Circles
X Lines, Contours
X Orientations
X Corners
X HOG, SIFT,…

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 36

Deep Learning – Features

X Hierarchical features are determined by deep learning

X No hand-designed features needed
X Example:

Mapping
Features Features Features
Input from Output
level 1 level 2 level 3
features

Source: https://www.nature.com/news/computer-science-the-
Auto.-Sys: Deep Learning learning-machines-1.14481, Images Andrew Ng, 27.08.2017 Prof. Dr. N. Stache 37

Observations: Hand-Designed Features

vs. Learned Representations

X Hand designed features: easy to understand

X In some cases: hand-designed features are difficult to define,
e.g. if variation range is high Æ learned representations
X Decision graph:
no
Deep learning
no yes
Classic approaches,
yes yes, no e.g. machine learning
with hand designed features

Formal description of Precise and intuitive

problem is understanding of features
easy to find? and mapping is vital?

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 38

 What is Deep Learning?
X Deep Learning is an approach for…
X generating hierarchical representations from
input data by itself and
X mapping these to an output.

When is deep learning beneficial?

X When formal description is difficult
X A precise understanding of features and the mapping to the output is
not critical
X But: Visualization of learned features can be valuable

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 39

Motivation for Deep Learning…

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 40

Feature Hierarchy Æ Bio-Inspired (!)

X Multiple stages in recognition pathway of mammals:

[picture from Simon Thorpe]

Auto.-Sys: Deep Learning Source: http://monkeylogic.uchicago.edu/old/Science.htm, 27.08.2017 Prof. Dr. N. Stache 41

ImageNet Large Scale Visual Recognition

Challenge (ILSVRC)
X ImageNet:
X > 14 000 000 annotated images
X > 20 000 classes
Classification:

X Goal: 1000 images per class

X ILSVRC:
X Classification
(i.e. what is in the image?
Æ 5 guesses + confidence)
X Classification with localization
(i.e. like classification + 1 bounding Source: https://blog.acolyer.org/2016/04/20/imagenet-classification-with-
deep-convolutional-neural-networks/, 27.08.2016
box for each guess)
X Detection
Detection:

(i.e. like classification with localization

but any number of objects possible,
false positives are penalized)

Source: http://image-net.org/challenges/LSVRC/2014/, 27.08.2016

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 42

ILSVRC Classification Top-5 Error (%)

classic machine learning deep learning

30 28,2 300

25,8 269
25 250

20 200

Number of layers
16,4
Error (%)

15 152 150

10 100
7,3 6,7

Human 5 3,6 3
50
19 22
8
0 0
2010 2011 2012 2013 2014 2015 2016
error (%) layers

Î Great Success of Deep Learning in ImageNet Challenge!

Source: http://image-net.org/challenges/talks_2017/ILSVRC2017_overview.pdf,
Auto.-Sys: Deep Learning 05.02.2017
Prof. Dr. N. Stache 43

Engineering Speedup by Deep Learning

X No hand-engineered features with deep learning

X Re-use of pretrained networks (Transfer-learning)

X George Hotz, 26-year old hacker creates self-driving car!?

Auto.-Sys: Deep Learning Source: https://www.theguardian.com/technology/2016/apr/05/george-hotz-comma-self- Prof. Dr. N. Stache 44

driving-car-tesla-elon-musk, 01.09.2017
Deep Learning is Hot in…

X Speech recognition
X Translation of spoken language into
nto text

X Natural Language Processing

X Understanding of natural language
e to perform tasks, e.g.
X Question answering
X Text summarization
X Search
X Machine translation
X Sentiment analysis

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 45

Deep Learning is Hot in…

X Computer vision
X Handwriting recognition
X Face & People Detection
X (Low-resolution) Object recognition
X Semantic labelling

X Applied mathematics

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 46

Automotive-Related Applications

X Detection of objects
X Detection of lane markings
X Detection of free space +
classification of the boundary

Source: https://www.youtube.com/watch?v=URmxzxYlmtg, 31.08.2017,

Talk by NVIDIA CEO Jen-Hsun Huang at GTC Europe

Source: https://www.youtube.com/watch?v=URmxzxYlmtg, 31.08.2017,

Source: https://www.youtube.com/watch?v=URmxzxYlmtg, 31.08.2017,
Talk by NVIDIA CEO Jen-Hsun Huang at GTC Europe
Talk by NVIDIA CEO Jen-Hsun Huang at GTC Europe

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 47

Automotive-Related Applications

X Image Segmentation

Auto.-Sys: Deep Learning Source: Synthia-Dataset, Prof. Dr. N. Stache 48

Bachelor Thesis Pascal Graf
 Automotive-Related Applications

X End-to End-Learning (e.g. NVIDIA BB8)

Source: https://www.youtube.com/watch?v=URmxzxYlmtg, 31.08.2017,

Talk by NVIDIA CEO Jen-Hsun Huang at GTC Europe

Visualization of
relevant image
regions for automated
driving:

Source: https://www.youtube.com/watch?v=URmxzxYlmtg, 31.08.2017,

Auto.-Sys: Deep Learning Talk by NVIDIA CEO Jen-Hsun Huang at GTC Europe
Prof. Dr. N. Stache 49

Why Deep Learning?

X Hierarchical features like in our brains
X Outperformed classical machine learning approaches
in mentioned areas
X Reduces development time
X Learn more about the (complex) problem by
visualizing deep learning features
X Seems to be successful in automotive applications

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 50

Why now? – Reasons for the Breakthrough

X Massive labeled datasets

X Massive parallel computing power
(e.g. GPUs)
X Software & algorithmic improvements
X Improved optimizers
– e.g. ADAM optimizer
X New regularization techniques
– e.g. dropout
X Backprop-friendly activation functions

simple Way to Prevent Neural Networks

Source: Srivastava, Hinton, Krizhevsky,
– e.g. ReLU

from Overfitting, J of Machine learning

sutskever, Salakhutdinov: Dropout: A

Research 15 (2014) 1929-1958.

X Software-Libraries
– e.g. TensorFlow, Theano, Caffe…

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 51

Fun Projects

X Style transfer
X Under Linux - open terminal, enter:
X conda create -n style-transfer1 python=3.5
X source activate style-transfer1
X pip install tensorflow
X conda install scipy pillow
X conda install -c conda-forge moviepy
X Under Windows – open terminal, enter:
X conda create -n style-transfer python=3.5
X activate style-transfer
X pip install tensorflow
X conda install scipy pillow
X conda install -c conda-forge moviepy

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 52

Fun Projects

X Style transfer
X Download the zip archive of fast style transfer and extract it
https://github.com/lengstrom/fast-style-transfer
X Download the Rain Princess checkpoint (model with already
tuned parameters), put it into the style transfer folder
https://goo.gl/niFGRc
X Copy your image into the fast style transfer folder
X Enter in your terminal:
X python evaluate.py --checkpoint ./rain-
princess.ckpt --in-path <path_to_input_file> --
out-path ./output_image.jpg

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 53

Fun Projects

My test image: Lake Constance

X Style transfer
X Other checkpoints:
X Rain-Princess: https://goo.gl/niFGRc
X La-muse: https://goo.gl/AqzB6P
X Udnie: https://goo.gl/B8gvAH
X Scream: https://goo.gl/dkA2K9
X Wave: https://goo.gl/UDTr3e
X Wreck: https://goo.gl/wRJh2v

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 54

Fun Project

X Deep Traffic
X Open: https://selfdrivingcars.mit.edu/deeptraffic/
X Modify the parameters
X lanesSide
X patchesAhead
X patchesBehind
X Number of hidden layers
X num_neurons
X Try to get it faster than 65 mph

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 55

Next week:

Python + Tooling + how to install

Î Please bring your Laptops!

Auto.-Sys: Deep Learning Prof. Dr. N. Stache 56