Ar#ficial Intelligence Lecture in the scope of the Honors Program
Quetelet Colleges March 14, 2017
Aleksandra Pizurica
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
² AI around us: recent progress
² Measuring intelligence: Beyond Turing test
² Can we trust robots? AI for the people
² Deep learning: Principles and state-‐of-‐the-‐art
² Some of UGent research: ApplicaBons in art invesBgaBon
2
Overview
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 3
AI entering all spheres of our life
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 4
High demand for AI professionals
IEEE The InsBtute, vol. 40, no. 2, June 2016
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
5
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
natural language processing – computer vision – robotics
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
6
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
game tree
route planning tree search
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
7
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
planning graphs
acBon schema
knowledge diagrams
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
8
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
belief propagaBon
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
9
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS) decision networks
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
10
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Main parts of modern AI
11
Problem solving Game playing (SEARCHING)
Knowledge representaBon
Planning (LOGIC)
Reasoning under uncertainty (BAYESIAN NETWORKS)
RaBonal decisions & AcBng (PROBABILITY + UTILITY)
Learning (NEURAL & BELIEF
DEEP NETS)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
The foundaBons of AI
12
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Autonomous smart vehicles • Driverless cars (e.g. Google car) • “highway pilots” for hands-‐free driving • PredicBons: 75% autonomous cars by 2040
State of the art in AI
13
A. Davies: AI Is All Around Us, IEEE The Ins1tute, June 2016.
Examples: • Released in 2016: BMW 750i xDrive
can park itself with no one behind the wheel
• 2015 Infinity Q50S and 2015 Mercedes-‐Benz S65 AMG engaging the breaks when car comes close to another object or pedestrian
• Jan 2016 Toyota releases plans to invest $50 million in AI program
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
State of the art in AI
14
O. Levander, IEEE Spectrum, February 2017
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Autonomous planning and scheduling in space explora#on § Beginnings: NASA’s remote agent program (1999) § Lunar Atmosphere and Dust Environment Explorer (LADEE), 2013 § Mars exploraBon (two NASA’s rovers landed on Mars in 2014); Aurora launch scheduled in 2018 (ESA).
State of the art in AI
15
LADEE approaching Lunar orbit ArBst’s concept (hkps://www.nasa.gov/)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Computer aided diagnosis § E.g. in radiology the computer output is already rouBnely used as a "second opinion" in assisBng radiologists' image interpretaBons
RoboBc surgery
State of the art in AI
16
Da Vinci surgery robot, designed by Xi Spine, eye, hip & knee, cancer & tumor operaBons
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Go game
17
• A strategy board game, originaBng from ancient China (more than 5500 years old).
• Played on 19x19 grid of lines with pieces called stones.
• The aim: surround more territory than the opponent. The strategy includes akacking the opponent's weak groups.
• Simple rules, but highly complex game, far more complex than chess! (b ≈ 250, d ≈ 150)
Source: Wikipedia hkps://en.wikipedia.org/wiki/Go_(game)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
AlphaGo
18
AlphaGo is a computer program developed by Google DeepMind
• In October 2015, the first Computer Go program that defeated a professional human Go player
• January 2016: defeated European Go champion Fan Hui (2 dan master) 5:0
• March 2016: AlphaGo defeated a 9-‐dan master Lee Sedol, 4:1
The method was published in Nature, Jan 2016 issue:
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
AlphaGo vs. human professional
19
D. Silver et al: Mastering the game of Go with deep neural networks and tree search, Nature, January 2016.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Some recently featured topics
20
From IEEE Computa1onal Intelligence Magazine, November 2014
Heterogeneous vehicle rouBng
OpBmizing supply chain networks
Planning of aircrar trajectories
…
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: Asteroid exploraBon
21
From IEEE Computa1onal Intelligence Magazine, October 2013
Autonomous Asteroid ExploraBon by RaBonal Agents
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: Rover missions
22
From IEEE Computa1onal Intelligence Magazine, October 2013
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: Smart Grid
23
From IEEE Computa1onal Intelligence Magazine, August 2011
ComputaBonal Intelligence for the Smart Grid
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: Smart Internet of Things
24
From IEEE Computa1onal Intelligence Magazine, August 2013
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: EmoBon analysis
25
Communica1ons of the ACM, December 2014
ComputaBonally Modeling Human EmoBon
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
ApplicaBons: Stock market predicBon
26
A. HedayaB Moghaddama, M. H. Moghaddamb, M. Esfandyari. Stock market index predicBon using arBficial neural network, Journal of Economics, Finance and Administra1ve Science, 21 (2016) 89–93.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Some recently featured topics
27
Intelligence Technology for Robots That Think From IEEE Computa1onal Intelligence Magazine, August 2013
Measuring Intelligence Beyond the Turing test
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
What is AI?
29
Four categories of AI definiBons
Systems that think like humans Systems that think ra#onally
Systems that act like humans Systems that act raBonally
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Intelligent Agents
30
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
The Turing test (Alan Turing, 1950) was designed to provide a saBsfactory operaBonal definiBon of intelligence
Suggested major components of AI: § Knowledge representaBon (store what it hears or knows) § Automated reasoning (use the stored info to draw conclusions) § Machine learning (adapt to new scenarios; detect and extrapolate pakerns) § Language processing (e.g., communicate in English or another language)
Extension -‐ total Turing test includes video to test perceptual abiliBes
The Turing test
31
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Importance of the Turing test
32
The experts in AI do not give much importance to actually passing the Turing test. Rather, its main importance is in defining the major components of the field
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Beyond the Turing test
33
TesBng math and geometry
TesBng commonsense knowledge
TesBng inference and world knowledge
AI Magazine, Spring 2016 issue
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Beyond the Turing test
34
PotenBals of AI systems: • GeneraBng and verifying quickly huge numbers of plausible hypotheses • Maintaining global repository of knowledge (access to huge amounts of
papers, experiments, reports etc.)
Discovery as a search problem: Deep exploraBon of knowledge space
AI Magazine, Spring 2016 issue
AI for the people Can we trust robots?
35
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Some recently featured topics
36
AIXI – An opBmal agent model for maximizing an environmental reward signal
Communica1ons of the ACM, September 2014 Exploratory Engineering in ArBficial Intelligence
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Some recently featured topics
37
IEEE Spectrum, June 2016
IEEE The Ins1tute, June 2016 Communica1ons of the ACM, September 2016
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
AI for rather than instead of people
38
PrioriBes for robust and beneficial AI: • OpBmizing AI’s economic impact • Law and ethics research
§ Liability & law for autonomous vehicles § Machine Ethics § Autonomous Weapons § Privacy
Robust and Beneficial ArBficial Intelligence: “Because of the great poten1al of AI, it is important to research how to reap its benefits while avoiding poten1al piQalls”
Interest in human-‐centered approach: • RehabilitaBon • Assist with disabiliBes • Facilitate learning
AI Magazine, Winter 2015
Deep learning Principles and state-‐of-‐the-‐art
39
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Example: Image classificaBon
40
Example from: Stéphane Mallat, "Scakering Invariant Deep Networks for ClassificaBon” (Caltech database)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Challenges in object classificaBon
41
Samoyed White wolf
At a pixel level, images of two samoyeds can be quite different depending on the pose and background, whereas images of a wolf and a samoyed can appear similar
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
A long standing dream of pakern recogniBon was: replace hand-‐engineered features with representa1on learning
Learning vs. feature engineering
42
Deep Learning: Yann LeCun, Yoshua Bengio & Geoffrey Hinton
Finding the right features is difficult, requires expert knowledge, tuning …
Classical machine learning
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
RepresentaBon learning § Feed machine with raw data § AutomaBcally discover representaBons
Deep learning § RepresentaBon learning with mulBple layers § Simple but non-‐linear modules at each level § Results in a hierarchy of representaBons
Key ideas § Layers of features not designed by programmers § Learning features from data with a general method
RepresentaBon learning
43
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Image features
44
LeCun-‐Bengio-‐Hinton, Deep Learning, Nature 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Image features
45
LeCun-‐Bengio-‐Hinton, Deep Learning, Nature 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Image features
46
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Why mulBple layers?
47
A mulBlayer neural network can distort the input space to make the classes of data linearly separable.
C. Olah (hkp://colah.github.io/)
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Supervised learning
48
NVIDIA GPUs -‐ The Engine of Deep Learning (hkps://developer.nvidia.com/deep-‐learning)
In a typical deep-‐learning system there may be hundreds of millions of weights and hundreds of millions of labeled examples with which to train a machine
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
OpBmizaBon is highly non-‐convex
49
J. Shlens and G. Toderici, Deep Learning for Image and Video Processing, ICIP 2016 tutorial
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Feed forward pass
50
At each layer: weighted sum of inputs followed by a nonlinearity. Typically, recBfied linear unit (ReLU) is used used:
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
BackpropagaBon
51
Error derivaBves propagate from top to to bokom by applying a simple rules
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Deep learning with CNN
52
ConvoluBonal Neural Networks (CNN) are biologically-‐inspired variants of mulB-‐layer percepBons
LeCun-‐Bengio-‐Hinton, Deep Learning, Nature 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
InspiraBon from neuroscience
53
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Feedforward model contd.
54
V1 -‐ primary visual cortex; V2 -‐ visual area II; V4 -‐ visual area IV; PIT -‐ posterio inferotemporal cortex; AIT -‐ anterior inferotemporal cortex;
Si -‐ simple cells at layer Vi Ci -‐ complex cells at layer Vi
T. Serre et al : A Theory of Object RecogniBon: ComputaBons and Circuits in the Feedforward Path of the Ventral Stream in Primate Visual Cortex
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Feedforward model
55
T. Poggio and E. Bizzi, GeneralizaBon in vision and motor control, Nature, 2004.
• Valid for the rapid categorizaBon tasks
• It is believed that first 100-‐200 milliseconds of visual percepBon involves mainly feedforward processing
• Human observers can dicriminate a scene that contains a parBcular prominent object (e.g., animal, vehicle) arer only 20ms of exposure
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
From images to text
56
Vision Deep CNN
Language generaBng RNN
RNN – Recurrent neural network (its hidden units maintain a ‘state vector’ that implicitly contains informaBon about the history of the inputs)
LeCun-‐Bengio-‐Hinton, Deep Learning, Nature 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Deep learning with CNN
57
A women is throwing a frisbee in a park A dog is standing on a hardwood floor A stop sign is on a road with a mountain in the background
A likle girl is si~ng on a bed with a teddy bear
A group of people si~ng on a boat in the water
A giraffe standing in a forest with trees in the background
LeCun-‐Bengio-‐Hinton, Deep Learning, Nature 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
State-‐of-‐the-‐art in deep learning
58
Some of our research ApplicaBons to art invesBgaBon
59
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 60
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 61
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Ghent Altarpiece: RestoraBon treatment
62
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 63
Ghent Altarpiece: RestoraBon treatment
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
AutomaBc detecBon of paint losses
64
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence 65
Virtual restoraBon
InpainBng method from: T. Ruzic and A. Pizurica, IEEE Transac1ons on Image Processing, 2015.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Virtual restoraBon
66
A. Pizurica, L. Platisa, T. Ruzic, et al. (2015): Digital Image Processing of the Ghent Altarpiece: Supporting the Painting’s Study and Conservation Treatmant. Signal Processing Magazine, 9(2): 583-594.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Digital painBng analysis D1
D2
D3
C1 C2
C3
B1
B2
A3
A2
A1
A4
Crack detecBon and virtual restoraBon
Painter style characterizaBon
67
A. Pižurica, L. Platisa, T. Ruzic, et al. (2015): Digital Image Processing of the Ghent Altarpiece: Supporting the Painting’s Study and Conservation Treatmant. Signal Processing Magazine, 9(2): 583-594.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
Other high-‐dimensional data Hyperspectral images
• Hundreds of spectral bands à huge data sets!
• Much richer informaBon, but a huge challenge for processing
Extract most interesting information from massive data
68
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
MulBmodal data fusion
Best Paper Challenge award: 2014 IEEE GRSS Data Fusion Contest
visible RGB
69
W. Liao, F. Van Coillie, A. Pizurica, S. Gautama and W. Philips (2014): Fusion of thermal infrared hyperspectral and VIS RGB data using guided filtering and supervised fusion graph, IGARSS’14.
A. Pizurica, Quetelet Colleges 2017 : ArBficial Intelligence
• Understanding the mechanisms behind deep learning § The number of layers, input and output neurons and filter responses are determined through experiments that require expert knowledge
§ Rigorous mathemaBcal models are sBll lacking
• Unsupervised learning § There are much more unlabeled data; humans learn by observing too
• CombinaBon of learning and complex reasoning § Includes efficient inference methods in graphical models
• PrioriBes for beneficial AI -‐ building AI for the people § OpBmizing AI’s economic impact § Law & ethics research
70
Some future prospects of AI research