Transcript Aprendizaje Automático en Astrofísica, Óptica y Otras

Artificial Intelligence
Past, Present, and Future
Olac Fuentes
Associate Professor
Computer Science Department
UTEP
1
Artificial Intelligence
A definition:
• AI is the science and engineering of making
intelligent machines
2
Artificial Intelligence
A definition:
• AI is the science and engineering of making
intelligent machines
But, what is intelligence?
• A very general mental capability that, among other
things, involves the ability to reason, plan, solve
problems, think abstractly, comprehend complex
ideas, learn quickly, and learn from experience.
3
Artificial Intelligence
Another definition:
• AI is the science and engineering of making
machines that are capable of:
–
–
–
–
–
–
Reasoning
Representing knowledge
Planning
Learning
Understanding (human) languages
Understanding their environment
4
Artificial Intelligence
• Weak AI Claim - Machines can possibly act as if
they were intelligent
• Strong AI Claim - Machines can actually think
intelligently
5
Artificial Intelligence
Why?
6
Artificial Intelligence
Why?
– Building an intelligent machine will help
us better understand natural intelligence
7
Artificial Intelligence
Why?
– Building an intelligent machine will help
us better understand natural intelligence
– An intelligent machine can be used to
perform difficult and useful tasks whether it models human intelligence or
not
8
Artificial Intelligence
Brief History
– Field was founded in 1956, initially led by John Mc
Carthy, Marvin Minsky, Allen Newell and Herbert
Simon (known as the “fourfathers” of A.I.)
– Great initial optimism, grandiose objectives
(“machines will be capable, within twenty years, of
doing any work a man can do” – H. Simon)
– Emphasis on symbolic reasoning
– Huge government spending
– Disappointing results
9
Artificial Intelligence - Brief History
The A.I. Winter
– 1966: the failure of machine translation,
– 1970: the abandonment of connectionism,
– 1971−75: DARPA's frustration with the Speech Understanding
Research program at Carnegie Mellon University,
– 1973: the large decrease in AI research in the United Kingdom in
response to the Lighthill report,
– 1973−74: DARPA's cutbacks to academic AI research in general,
– 1987: the collapse of the Lisp machine market,
– 1988: the cancellation of new spending on AI by the Strategic
Computing Initiative,
– 1993: expert systems slowly reaching the bottom,
– 1990s: the quiet disappearance of the fifth-generation computer
project's original goals,
10
Artificial Intelligence
Brief History – The Comeback
– Rebirth of Connectionism
• The backpropagation algorithm (Hinton and others, 1986, PDP group)
– Machine learning becomes usable
• The ID3 and C4.5 algorithms – decision trees for the masses - R. Quinlan, 86
• Increased computing power
• Increased availability of data in electronic form
– Behavior-based (or “emerging”) A.I.
• “A robust layered control system for a mobile robot” – R. Brooks, 85
• “Intelligence is in the eye of the beholder”, “The world is its own best model”,
“Elephants don’t play chess”, “We don’t need no representation”
• Agent-based architectures (Maes, and many others)
– Active Vision
• The goal of machine perception is not to build a 3D model of the world, but to
extract information to perform useful tasks (D. Ballard, Y. Aloimonos)
11
Artificial Intelligence
Brief History – Present Times
–Realistic expectations
–Lots of useful applications
–Research divided into subareas
(vision, learning, NLP, planning,
etc.)
–Little work on overall
intelligence
12
Artificial Intelligence
The Old Times
The pursuit of “General AI”
Objective: Build a machine that exhibits ALL
of the AI features
13
Old Times – The Turing Test
How do we know when AI research has
succeed?
When a program that can consistently pass the
Turing test is written.
14
Old Times – The Turing Test
A human judge engages in a natural
language conversation with one
human and one machine, each of
which tries to appear human; if the
judge cannot reliably tell which is
which, then the machine is said to
pass the test.
15
Old Times – The Turing Test
Problems with the Turing test:
16
Old Times – The Turing Test
Problems with the Turing test:
• Human intelligence vs. general intelligence
– Computer is expected to exhibit undesirable
human behaviors
– Computer may fail for being too smart
17
Old Times – The Turing Test
Problems with the Turing test:
• Human intelligence vs. general intelligence
– Computer is expected to exhibit undesirable
human behaviors
– Computer may fail for being too smart
• Real intelligence vs. simulated intelligence
18
Old Times – The Turing Test
Problems with the Turing test:
• Human intelligence vs. general intelligence
– Computer is expected to exhibit undesirable
human behaviors
– Computer may fail for being too smart
• Real intelligence vs. simulated intelligence
• Do we really need a machine that passes it?
19
Old Times – The Turing Test
Problems with the Turing test:
• Human intelligence vs. general intelligence
– Computer is expected to exhibit undesirable
human behaviors
– Computer may fail for being too smart
• Real intelligence vs. simulated intelligence
• Do we really need a machine that passes it?
• Testing the machine vs. testing the judge
20
Old Times – The Turing Test
Problems with the Turing test:
• Human intelligence vs. general intelligence
– Computer is expected to exhibit undesirable
human behaviors
– Computer may fail for being too smart
•
•
•
•
Real intelligence vs. simulated intelligence
Do we really need a machine that passes it?
Testing the machine vs. testing the judge
Too hard! – Very useful applications can be
built that don’t pass the Turing test
21
More Recent Research
Goal: Build “intelligent” programs that are useful for a
particular task
Normally restricted to one target intelligent behavior.
Thus AI has been broken into several sub-areas:
– Machine learning
– Robotics
– Computer vision
– Natural language processing
– Knowledge representation and reasoning
22
What has AI done for us?
State of the Art
It has provided computers that are able to:
• Learn (some simple concepts and tasks)
• Allow robots to navigate autonomously (in
simplified environments)
• Understand images (of restricted predefined types)
• Understand human languages (some of them,
mostly written, with limited vocabularies)
• Reason (using brute force, in very restricted
domains)
23
What has AI done for us?
Machine Learning – Netflix movie recommender system
Very active research area
–
–
–
–
–
–
Extract statistical regularities from data
Find decision boundaries
Find decision rules
Imitate human brain
Imitate biological evolution
Combine several approaches
24
What has AI done for us?
Machine Learning – Netflix movie recommender system
Idea:
• After returning a movie, user assigns a grade to it
(from 1 to 5)
• Given (millions) of records of users, movies and
grades, and the pattern of grades assigned by the
user, the system presents a list of movies the user
is likely to grade highly
25
What has AI done for us?
Robotics - Stanley, a self-driving car
26
What has AI done for us?
Robotics - Stanley, a self-driving car
What does Stanley learn?
A mapping from sensory inputs to driving commands
27
What has AI done for us?
Robotics - Lexus self-parking system
28
What has AI done for us?
Computer Vision - Face Detecting
Cameras
29
What has AI done for us?
Computer Vision - Face
Detecting Cameras
30
What has AI done for us?
Reasoning
Successful applications:
• Route planning systems
• Game playing programs
31
What has AI done for us?
Reasoning
The Zohirushi Neuro Fuzzy® Rice Cooker & Warmer features advanced Neuro
Fuzzy® logic technology, which allows the rice cooker to 'think' for itself and
make fine adjustments to temperature and heating time to cook perfect rice
every time.
32
What has AI done for us?
Natural language processing
Successful applications:
• Dictation systems
• Text-to-speech systems
• Text classification
• Automated summarization
• Automated translation
33
What has AI done for us?
Natural language processing
Automated Translation
Original English Text:
The Dodgers became the fifth team in modern major league
history to win a game in which they didn't get a hit,
defeating the Angels 1-0. Weaver's error on a slow roller led
to an unearned run by the Dodgers in the fifth.
What has AI done for us?
Natural language processing
Automated Translation
Original English Text:
The Dodgers became the fifth team in modern major league
history to win a game in which they didn't get a hit,
defeating the Angels 1-0. Weaver's error on a slow roller led
to an unearned run by the Dodgers in the fifth.
Translation to Spanish (by Google - 2009)
Los Dodgers se convirtió en el quinto equipo en la moderna
historia de las ligas mayores para ganar un juego en el que
no obtener una respuesta positiva, derrotando a los Ángeles
1-0. Weaver's error en un lento rodillo dado lugar a un
descontados no correr por la Dodgers en el quinto.
What has AI done for us?
Natural language processing
Automated Translation
Translation to Spanish (by Google - 2009)
Los Dodgers se convirtió en el quinto equipo en la moderna
historia de las ligas mayores para ganar un juego en el que
no obtener una respuesta positiva, derrotando a los Ángeles
1-0. Weaver's error en un lento rodillo dado lugar a un
descontados no correr por la Dodgers en el quinto.
What has AI done for us?
Natural language processing
Automated Translation
Translation to Spanish (by Google – 2009)
Los Dodgers se convirtió en el quinto equipo en la moderna
historia de las ligas mayores para ganar un juego en el que
no obtener una respuesta positiva, derrotando a los Ángeles
1-0. Weaver's error en un lento rodillo dado lugar a un
descontados no correr por la Dodgers en el quinto.
Translation back to English (by Google – 2009)
The Dodgers became the fifth equipment in the modern history
of the leagues majors to gain a game in which not to obtain
a positive answer, defeating to Los Angeles 1-0. Weaver' s
error in a slow given rise roller to discounting not to run by
the Dodgers in fifth.
What has AI done for us?
Natural language processing
Automated Translation
Original English Text:
The Dodgers became the fifth team in modern major league
history to win a game in which they didn't get a hit,
defeating the Angels 1-0. Weaver's error on a slow roller led
to an unearned run by the Dodgers in the fifth.
What has AI done for us?
Natural language processing
Automated Translation
Original English Text:
The Dodgers became the fifth team in modern major league
history to win a game in which they didn't get a hit,
defeating the Angels 1-0. Weaver's error on a slow roller led
to an unearned run by the Dodgers in the fifth.
Translation to Spanish (by Google - 2010)
Los Dodgers se convirtió en el quinto equipo en la historia
moderna de Grandes Ligas en ganar un juego en el que no
recibieron una respuesta positiva, derrotando a los
Angelinos 1-0. error de Weaver en una rola lenta dio lugar a
una carrera sucia por los Dodgers en el quinto.
What has AI done for us?
Natural language processing
Automated Translation
Translation to Spanish (by Google – 2010)
Los Dodgers se convirtió en el quinto equipo en la historia
moderna de Grandes Ligas en ganar un juego en el que no
recibieron una respuesta positiva, derrotando a los
Angelinos 1-0. error de Weaver en una rola lenta dio lugar a
una carrera sucia por los Dodgers en el quinto.
Translation back to English (by Google – 2010)
The Dodgers became the fifth side in the modern history of
baseball to win a game that did not get a hit, defeating the
Angels 1-0. Weaver's error on a slow roller led to an
unearned run for the Dodgers in the fifth.
The Future of AI
41
The Future of AI
Making predictions is hard, especially about the future - Yogi
Berra
42
The Future of AI
Making predictions is hard, especially about the future - Yogi
Berra
But…
• Continued progress expected
• Greater complexity and autonomy
• New enabling technology - Metalearning
• Once human-level intelligence is attained, it will be quickly
surpassed
43
Conclusions
44
Conclusions
• Artificial Intelligence has made a great deal of progress
since its inception in the 1950s
45
Conclusions
• Artificial Intelligence has made a great deal of progress
since its inception in the 1950s
• The goal of general AI has been abandoned (at least
temporarily)
46
Conclusions
• Artificial Intelligence has made a great deal of progress
since its inception in the 1950s
• The goal of general AI has been abandoned (at least
temporarily)
• Useful applications have appeared in all subfields of AI,
including: Machine learning, computer vision, robotics,
natural language processing and knowledge representation
47
Conclusions
• Artificial Intelligence has made a great deal of progress
since its inception in the 1950s
• The goal of general AI has been abandoned (at least
temporarily)
• Useful applications have appeared in all subfields of AI,
including: Machine learning, computer vision, robotics,
natural language processing and knowledge representation
• The field continues to evolve rapidly
48
Conclusions
• Artificial Intelligence has made a great deal of progress
since its inception in the 1950s
• The goal of general AI has been abandoned (at least
temporarily)
• Useful applications have appeared in all subfields of AI,
including: Machine learning, computer vision, robotics,
natural language processing and knowledge representation
• The field continues to evolve rapidly
• Increased complexity and unpredictability of AI programs
will raise important ethics issues and concerns
49
AI and Psychology
Some questions/issues:
• Can A.I. algorithms be used to model natural
intelligence?
• How can we exploit our knowledge of human
intelligence to develop artificially intelligent
systems?
• Can psychology help settle the Strong A.I. vs.
Weak A.I. debate?
50
THANKS!
Questions?
For more info:
http://www.cs.utep.edu/ofuentes/
51
UTEP’s Vision and Learning
Laboratory
Our goals:
Programming computers to see
Programming computers to learn
52
UTEP’s Vision and Learning
Laboratory
A Sample of Current Research Projects
•
•
•
•
•
•
•
•
Transfer learning using deep neural networks
Image super-resolution
Image compression
Tracking multiple nearly-identical objects
Vision with foveal cameras
Medical image analysis
Astronomical data analysis
Predicting RNA folding
53
Machine Learning
Why?
Scientific Reason:
• Learning is arguably the most important feature of
intelligence.
• If we want to understand and replicate intelligent
behavior, the ability to learn is indispensable.
54
Machine Learning
Why?
Engineering Reason:
• Programs that allow computers to learn are the
only practical way of solving a wide variety of
very difficult problems.
55
Machine Learning
The key enabling technology of AI
Problem Solving in Computer Science
56
Machine Learning
The key enabling technology of AI
Problem Solving in Computer Science
• Traditional Approach
– Write a detailed sequence of instructions (a program)
that tells the computer how to solve the problem.
57
Machine Learning
The key enabling technology of AI
Problem Solving in Computer Science
• Traditional Approach
– Write a detailed sequence of instructions (a program)
that tells the computer how to solve the problem.
• Machine Learning Approach
– Give the computer examples of desired results and let it
learn how to solve the problem.
58
Machine Learning
The key enabling technology of AI
Problem Solving in Computer Science
• Traditional Approach
– Write a detailed sequence of instructions (a program)
that tells the computer how to solve the problem.
• Machine Learning Approach
– Give the computer examples of desired results and let it
learn how to solve the problem.
– Advantage: It allows to solve problems that we can’t
solve with the traditional approach
59
Machine Learning
The key enabling technology of AI
Problem Solving in Computer Science
• Traditional Approach
– Write a detailed sequence of instructions (a program)
that tells the computer how to solve the problem.
• Machine Learning Approach
– Give the computer examples of desired results and let it
learn how to solve the problem.
– Advantage: It allows to solve problems that we can’t
solve with the traditional approach
– Most applications in other AI areas are based on machine
learning
60
Background: Artificial Neural
Networks
Idea: Imitate the information processing
capabilities of the central nervous system using a
large set of simple parallel non-linear units
(“artificial neurons”)
61
Background: Artificial Neural
Networks
The output of a neuron is given by a combination of its inputs and
its associated weights
We use a set of examples to adjust the weights to allow the
network to give the correct output – this is called training the
network
We use the trained network to solve problems!
62
Artificial Neural Networks
Face Recognition
Input: Pixel intensities
Output: Vicente Fox!
Input: Pixel intensities
Output: Angelina Jolie!
63
Artificial Neural Networks
Classification of Infant Cry
Input: Sound intensities
over time
Output: Hungry
Input: Sound intensities
over time
Output: Pain
64
Artificial Neural Networks
Driving Autonomous Vehicles
Input: Pixel
intensities
Output: Left
Input: Pixel
intensities
Output: Straight
Input: Pixel
intensities
Output: Right
65
Background: Artificial Neural
Networks
Advantages:
• Can approximate many types of functions (real-valued,
discrete-valued, vector-valued)
• Relatively insensitive to noise
Disadvantages:
• Training times can be long
• Overfitting –network “memorizes” training data and is
unable to generalize
66
Transfer learning using deep
neural networks
Research Question: Can we take advantage of
previously-acquired knowledge in order to learn
to solve a new task more quickly and/or with less
training data?
It works for humans:
- Knowing Spanish helps to learn Portuguese
- A tennis player can easily learn to play squash
67
Transfer learning using deep
neural networks
Idea # 1:
• Use a network with many layers
• Train network for a task
• Retrain only the last one or two layers in order to learn a new but
similar task
Idea # 2:
•
•
•
•
Use output layer with many neurons
Assign randomly chosen values to output neurons for each class
Train network for a task
When new class is added to task, find average value in output layer
and assign that as the label for the class
68
Transfer learning using deep
neural networks
Idea # 1:
• Use a network with many layers
• Train network for a task
• Retrain only the last one or two layers in order to learn a new but
similar task
Results
• We trained a network to recognize handwritten letters
• We can re-train a network to recognize digits using less than 5% of
the time and training examples
Potential Applications
• Speech recognition
• Face recognition
69
Transfer learning using deep
neural networks
Idea # 2:
•
•
•
•
Use output layer with many neurons
Assign randomly chosen values to output neurons for each class
Train network for a task
When new class is added to task, find average value in output layer
and assign that as the label for the class
Results
• We trained a network to recognize handwritten letters
• Network can recognize digits without any retraining
Potential Applications
• Flexible face recognition systems
• Language models with unlimited vocabulary
70
Image Super-resolution
Research Question: Can we increase the resolution of an
image of a known class (say faces, or hand-written text)
in software? That is, given a low-resolution (LR) image
of an object, can we infer its appearance in highresolution (HR)?
Idea:
Crate many pairs of (LR,HR) images
Learn function from LR to HR (the opposite is trivial)
When given a LR image, apply learned function to increase
resolution
71
Image Super-resolution
LR input image
Generated image
Original HR image
72
Image Super-resolution
Theory and Algorithms
 Decouple high-resolution face image to two parts
=
+
I Hg
I Hl
I H — high resolution face image I Hg — global face I Hl — local face
IH
 Two-step Bayesian inference
I H*  arg max p ( I L | I H ) p ( I H )
IH
IH
?
 arg max p ( I L |
I Hg , I Hl
I Hg , I Hl ) p ( I Hg , I Hl )
 arg max p ( I L | I Hg ) p ( I Hg ) p ( I Hl | I Hg )
IL
I Hg , I Hl
1. Inferring global face
1. Inferring
global face
g*
I H  arg max p( I L | I Hg ) p( I Hg )
I Hg *
2. Inferring local face
2. Inferring
local face
I Hl *  arg max
p ( I Hl | I Hg * )
I Hl *
Finally adding them together
Finally
adding them together
I H*  I Hl *  I Hg *
73
Image Super-resolution
Our Contributions:
• Improved image alignment: k-nearest-neighbors warping
function
• More efficient reconstruction: a stochastic algorithm to
build local model
74
Image Super-resolution
Experimental results
(a)
(b)
(c)
(d)
(e)
(a) LR image Restored full face. (c) Restored global face. (d) Interpolated face. (e) Original face. (f) Restored local face.
(f)
75
Compression of Medical Images
using Super-resolution
Observation: if we can derive the HR image from the
LR one, then we don’t need to store the HR image!
Idea:
• Segment image into regions of high, medium and
low importance
• Compress low importance pixels to a single bit
• Compress/decompress medium importance pixels
using super-resolution
• Compress/decompress high importance area using
lossless compression algorithm such as jpeg2000 76
Compression of Medical Images
using Super-resolution
a) Original mammogram; b), c) and d) show different
compression/decompression results for different compression
parameters
77
Compression of Medical Images
using Super-resolution
Compression Method PSNR Compression Ratio
JPEG 2000
41.95
80:1
Lossless JPEG
infinite
3:1
Our method
35.90
20480:1
Experimental results
78
Tracking Nearly Identical
Objects
Problem:
Given a video with multiple moving objects
that are similar, determine the trajectories of
each individual object
79
Tracking Nearly Identical Objects
80
Tracking Nearly Identical Objects
Approach:
• Build probabilistic models of motion of objects from
training video
• Find objects on consecutive frames f(i) and f(i+1)
• Use a backtracking search algorithm to find the set
of matches that maximizes probability, given models
Application:
Tracking flies – experiments to determine the effect of
alcohol ingestion on groups of flies
81
Tracking Nearly Identical Objects
82
Tracking Nearly Identical Objects
83
Traffic Sign Inspection and Tracking
Tracking – Locating an object of interest in a sequence of
images, despite variations in size and appearance
84
Machine Learning in Science
• Problem: Gathering scientific data is easy, gaining knowledge
from them is not!
– Huge amounts of data gathered with automated devices – Digital sky
surveys, medical databases, particle physics experiments, seismic data
– Data easily available through the Internet
– Not enough scientists to fully exploit data
• Thus, there's a need for automated methods to classify and
analyze the data and to derive knowledge and insight from
them.
• Machine learning offers a very promising methodology to attain
these goals
85
Example: Estimation of Stellar
Atmospheric Parameters from Spectra
What's a spectrum?
A plot of energy flux against wavelength. It shows the combination of
black body radiation (originated in the core) and absorption lines
(originated in the atmosphere)
I
86
Estimation of Stellar Atmospheric
Parameters
• Quantum theory tells us that an atom can absorb energy only
at certain discrete wavelengths.
• Thus, we can view an absorption line as the signature of the
atom, ion or molecule that produces it.
• An expert astronomer can estimate the properties of the star’s
atmosphere from the strength of the absorption lines.
• In particular, the effective temperature, surface gravity and
metallicity can be estimated from the strength of absorption
lines.
87
Estimation of Stellar Atmospheric
Parameters - Problem: Too many stars!
88
Experimental Results
Estimation of Stellar Atmospheric
Parameters
89
Experimental Results
Estimation of Stellar Atmospheric
Parameters
Ensemble
LWLR
OC
Error
Reduction
Teff[k]
143.33
126.88
11.5%
Log g[dex]
0.3221
0.2833
12%
Fe/H
0.223
0.172
22.9%
90
Vision with foveal cameras
Mission: transmit faces
28x37
chip
76x150
chip
Before transmission:
•Tune/enhance resolution
•Compress appropriately
46x35 full frame
Sampled from 3600x2700 sensor
91
Challenge – where to foveate?
Problem: Foveation is useful as long as we can effectively select the image regions where
super-resolution should be applied.
Our Approach:
Use motion as a first trigger of higher resolution
Learn from multiple positive and negative examples of objects of the class of interest
(e.g. faces, pedestrians, vehicles)
Challenges:
– The same algorithm must be applicable to widely varying object classes
– Choice of features
– Choice of (negative) examples
– Combination of detectors
Results:
•
We have built systems that detect over 90% of the faces in images, while yielding
a false-positive rate of about 0.00001%
•
Our methods can easily be implemented in systolic hardware
92
Foveating on faces
93
Foveating on faces
94
Motion-triggered Foveation
95
THANKS!
For more info:
http://www.cs.utep.edu/ofuentes/
96
Acknowledgements
This work was actually done by:
• Steven Gutstein
• Jason Zheng
• Geovany Ramirez
• Diego Aguirre
• Joel Quintana
• Peter Kelley
• Manali Chakraborty
• Trilce Estrada
97