Download Document

Document related concepts

Cognitive epidemiology wikipedia , lookup

Intelligence quotient wikipedia , lookup

Human intelligence wikipedia , lookup

Neuroscience and intelligence wikipedia , lookup

Person of Interest (TV series) wikipedia , lookup

Environment and intelligence wikipedia , lookup

Semantic Web wikipedia , lookup

Intelligence wikipedia , lookup

Evolution of human intelligence wikipedia , lookup

Artificial general intelligence wikipedia , lookup

Artificial intelligence wikipedia , lookup

Transcript
Why Machine Intelligence is
Very Hard
Theo Pavlidis
Distinguished Professor Emeritus
Dept. of Computer Science
[email protected]
http://theopavlidis.com
Limitations of Computers
• Some tasks (e.g. number factorization) are very hard
for computers (unless it is proven that NP = P), but
they are also very hard for humans. We do not
discuss such tasks in this talk.
• But there are also tasks that are quite easy for
humans but very hard for computers such as
language translation, image understanding, speech
recognition, game playing, etc. All these go often
under the name of Artificial Intelligence (AI).
Comparatively little progress has been in most of
them even though there is no theoretical reason for
their difficulty.
5/24/2017
Machine Intelligence - web version
2
The State of Machine Vision
• There have seen some successes, notably in
industrial inspection and reading of printed text but
a lot of problems remain open.
• Reading distorted text (CAPTCHA) is so hard that
it is used as a security device.
• Content Based Image Retrieval (CBIR) is
hopelessly behind content based text retrieval.
• Face recognition programs are known mainly for
their failure to perform outside the laboratory.
5/24/2017
Machine Intelligence - web version
3
CAPTCHA
• Completely
Automated
Public
Turing test to tell
Computers and
Humans
Apart
5/24/2017
Machine Intelligence - web version
4
Content-based Image Retrieval
(CBIR)
• Given an image find those that are similar to it
from a data base of images. (If the images
are labeled, the problem is reduced to text
search.)
• Many systems have been advertised but they
do well only on rather trivial queries.
• This should be contrasted with the success of
text retrieval, not only Google but earlier
programs such as the Unix grep.
5/24/2017
Machine Intelligence - web version
5
Example - 1
5/24/2017
Machine Intelligence - web version
6
Example - 2
5/24/2017
Machine Intelligence - web version
7
Why the Failures?
• More specifically why computers have so
much trouble with pictures (compared, say, to
alphanumeric data)?
5/24/2017
Machine Intelligence - web version
8
Human Intelligence made simple
Input
Concept
Input
Output
5/24/2017
Machine Intelligence - web version
9
The Big Difference
• The transformation of input to concept is a complex
process (binding), barely understood by
neuroscientists. (In spite of claims to the opposite by
some computer scientists.)
• It is hard to develop algorithms
for a barely understood process.
• Humans can transform concepts into formal entities
(words in a language) and then code them in
computer readable form.
• Computers can deal with such formal input.
5/24/2017
Machine Intelligence - web version
10
What Neuroscientist Say
• “Perceptions emerge as a result of
reverberations of signals between different
levels of the sensory hierarchy, indeed across
different senses”. The author then goes on to
criticize the view that “sensory processing
involves a one-way cascade of information
(processing)”
• Source: V.S. Ramachandran and S. Blakeslee Phantoms in the
Brain, William Morrow and Company Inc., New York, 1998 (p. 56)
5/24/2017
Machine Intelligence - web version
11
Reading Demo - 1
5/24/2017
Machine Intelligence - web version
12
Reading Demo - 1
Tentative binding on the letter shapes (bottom
up) is finalized once a word is recognized (top
down). Word shape and meaning over-ride early
cues.
5/24/2017
Machine Intelligence - web version
13
Reading Demo -2
New York State lacks proper facilities
for the mentally III.
The New York Jets won Superbowl III.
• Human readers may ignore entirely the shape of
individual letters if they can infer the meaning
through context.
5/24/2017
Machine Intelligence - web version
14
Reasons for the Poor Results in
Machine Vision and CBIR
• Images are represented by statistics of pixel
values (e.g. color histogram, texture
histogram, etc)
• Such statistics are unrelated to human
perception.
• Papers describing CBIR methods use trivial
queries (e.g. “show me all pictures with a lot
of green”).
5/24/2017
Machine Intelligence - web version
15
Perceptual versus Computational
Similarity
• Two pictures may differ a lot in their pixel
values but appear similar to a person. (“They
have the same meaning”.)
• Two pictures may differ in very few pixels but
they have different meaning. (Face portraits
of two different people in front of the same
background.)
5/24/2017
Machine Intelligence - web version
16
Perceptual versus Computational
Similarity
Perceptually close
5/24/2017
Pixel-wise close
Machine Intelligence - web version
17
Text versus Pictures
• In text files each byte (or two) is a numerical
code for a character. Therefore strings of
bytes correspond to words that carry
semantic meaning.
• In pictures each byte (or group thereof)
represents the color at a particular location
(pixel). Pixels are quite far from the
components that have a semantic meaning.
5/24/2017
Machine Intelligence - web version
18
We do not that well in text!
• If it is hard to search for concepts unless we
can map concepts into words.
• Example 1: Find all articles critical of the
government policy in dealing with the banking
crisis.
• Example 2: Find all articles about a dog
named Lucy. Amongst the Google returns
was an article with the phrase: “Lucy and I spent
the weekend alone together. We have a dog named
Kyler.”
5/24/2017
Machine Intelligence - web version
19
The Importance of Context
• “Human intelligence almost always thrives on
context while computers work on abstract
numbers alone. … Independence from
context is in fact a great strength of
mathematics.”
• Source: Arno Penzias Ideas and Information,
Norton, 1989, p. 49.
5/24/2017
Machine Intelligence - web version
20
Example of Using Context - 1
5/24/2017
Machine Intelligence - web version
21
Example of Using Context - 2
5/24/2017
Machine Intelligence - web version
22
Example of Using Context - 3
• A human needs to detect only part of the
contour of an object to recognize the object.
• It is wasteful to look for algorithms that will
produce complete contours without also
producing noise.
• The key to image analysis is to be able to
make decisions from incomplete data.
5/24/2017
Machine Intelligence - web version
23
Example of Using Context - 4
5/24/2017
Machine Intelligence - web version
24
The Challenges
• We need to replicate complex transformations
that the (human/animal) brain has evolved to do
over millions of years.
• We have to deal with the fact the processing is
not unidirectional and also affected by other
factors than the input (context). (Such factors
cause visual illusions.)
5/24/2017
Machine Intelligence - web version
25
A time scale
• The human visual system has evolved from
animal visual systems over a period of more
than 100 million years.
• Speech is barely over 100 thousand years old.
• Written text is no more than 10 thousand years
old.
5/24/2017
Machine Intelligence - web version
26
A note on brain models
• There is a history for considering the latest
technology to be a model of the human brain,
for example in the 16th century irrigations
networks were considered to be models of
the brain.
• If someone claims to have a machine
modeling the human brain, ask how could the
machine be modified to model the brain of a
dog (since a dog cannot learn to write poetry,
play chess, etc)?
5/24/2017
Machine Intelligence - web version
27
A Note on Neural Nets
Is this a model of the brain?
As much as a table is a model of a dog.
5/24/2017
Machine Intelligence - web version
28
Projections
• When we map pictures of, say, 1000 by 1000
pixels into an array of, say, 100 numbers we
perform, in effect, a projection from a high
dimensional space to one of lower
dimensions.
• If the number of samples is small we are
likely to be able to find structure in the
projection.
5/24/2017
Machine Intelligence - web version
29
Statistical Games
5/24/2017
Machine Intelligence - web version
30
Statistical Games
5/24/2017
Machine Intelligence - web version
31
Dead End Approaches
• “Training” on large numbers of samples has
been used as a way out of this problem.
• But humans (and animals) do not need to be
trained on large numbers of samples.
• Rats trained to distinguish between a square
and a rectangle perform quite well when
faced with skinnier rectangles. They have the
concept of rectangle!
5/24/2017
Machine Intelligence - web version
32
Distinguish Rectangles from Squares
The Artificially Intelligent Approach
• Take a hundred (or more) pictures of
rectangles and squares, compute several
statistics on each picture and for each picture
create a “feature” vector F. Then compute a
vector W so that
F’W > 0 for squares and
F’W < 0 for rectangles
5/24/2017
Machine Intelligence - web version
33
Distinguish Rectangles from Squares
The Natural Approach
• Find the outline of a shape (if one exists in a
picture) and fit a rectangle to it. Then
compute the aspect ratio of the rectangle. If it
is near 1 (for some given tolerance), then it is
called a square, otherwise a rectangle.
• Criticism: Method lacks generality!!!
5/24/2017
Machine Intelligence - web version
34
No Generality in Nature
• The animal visual systems has many special
areas for visual tasks (about 30 in the human
case).
• We have already seen examples where “high
level” (context) recognition takes quickly over
the low level data processing.
5/24/2017
Machine Intelligence - web version
35
The Learning Machine (neural net)
Approach
• It has the appeal of getting something for
nothing, so it is kept alive.
• We can “solve” a problem without really
understanding it.
• Give a learning machine “enough” samples
and a classifier will be found!!!
• (Forget about the rat who only needs two
samples.)
5/24/2017
Machine Intelligence - web version
36
Criteria for Choosing a
Problem to Work on
• Context should either be known or not important.
• Processing of the input should be relatively
simple (it should be clear what kind of information
we need to extract).
• For an example relying heavily on context see:
technology/BoxDimensions/overview.htm on my web
site.
• Comments on major areas in the next few slides.
5/24/2017
Machine Intelligence - web version
37
Speech Recognition
• Grammar driven models (using low level
context) have been quite successful.
• High level context is even better. For
example, matching a speech fragment to a
name on a list.
5/24/2017
Machine Intelligence - web version
38
Optical Character Recognition
(OCR)
• Printed text characters have small shape
variability and high contrast with the
background. (CAPTCHA systems negate
these properties)
• Spelling checkers (or ZIP code directories in
postal applications) introduce low level
context.
5/24/2017
Machine Intelligence - web version
39
An Aside: Why did OCR mature when
the need for it was diminished?
• The algorithms used in the products of the
1990s were known earlier but they were too
complex to be implemented effectively with
the digital technology of earlier times.
• When computer hardware became cheap
enough for good OCR, it also became cheap
enough for PCs and the Internet.
• Keep this in mind in your business plans!
5/24/2017
Machine Intelligence - web version
41
Face Recognition
• It took over forty years to built acceptable
quality machines that recognize written
symbols. What makes us think that we can
solve the much more complex problem of
distinguishing human faces?
• Neuroscientists point out that humans have
special neural circuitry for face recognition.
5/24/2017
Machine Intelligence - web version
42
Can you tell how these two pictures differ?
5/24/2017
Machine Intelligence - web version
43
How about these two?
5/24/2017
Machine Intelligence - web version
44
How these two faces differ?
5/24/2017
Machine Intelligence - web version
45
How about these two?
5/24/2017
Machine Intelligence - web version
46
How these two differ?
5/24/2017
Machine Intelligence - web version
47
How about these two?
5/24/2017
Machine Intelligence - web version
48
Face Recognition and
Scalability
• The population samples in published studies
are relatively small and include men and
women of different races with different
hairstyles, etc.
• I have never seen a study where all the
subjects are similar. For example, white blond
men between the ages of 20 and 30 with long
hair and beards.
• Subjects in published studies are
cooperative.
5/24/2017
Machine Intelligence - web version
49
Results from the Field
• Not surprisingly, the results of installed face
recognition systems have been dismal. An
ACLU press release of May 14, 2002 stated
that "interim results of a test of facerecognition surveillance technology … from
Palm Beach International Airport confirm
previous results showing that the technology
is ineffective."
• See also The Economist, October 26, 2002
5/24/2017
Machine Intelligence - web version
50
Face Detection
• Before proceeding with face recognition we
need to find the faces in a picture (face
detection)
• CMU has a web site where the public may
submit pictures and they get back results with
a green square overlaid on faces facing front
and green pentagons of profiles.
• Results are not robust.
5/24/2017
Machine Intelligence - web version
51
Glimpses from the Face
Detection Gallery - 1
5/24/2017
Machine Intelligence - web version
52
Glimpses from the Face
Detection Gallery - 2
No faces have been detected
5/24/2017
Machine Intelligence - web version
53
Glimpses from the Face
Detection Gallery - 3
They got the wrong person
5/24/2017
Machine Intelligence - web version
54
Concluding Remarks
• Before we attempt to built a machine to
achieve a goal we must ask ourselves
whether that goal is compatible with the laws
of nature (as we know them).
• While such laws are clear in Physics and
Chemistry, there are not in the field of
Computation except in some extreme cases.
5/24/2017
Machine Intelligence - web version
55
Human Credulity - 1
• In spite of well understood laws of physics
“inventors” persist in offering designs that
violate them and they find takers.
• Therefore fundamental advances in
Computer Science are likely to reduce but not
to eliminate preposterous claims.
5/24/2017
Machine Intelligence - web version
56
Human Credulity - 2
• 50 years ago Langmuir debunked UFOs but
also predicted that UFOs will be with us for a
long time because it is too good a story for
the news media to let go.
• The view of computers as giant brains that
are able to outthink and replace humans is
about as valid as visits by extraterrestrials,
but it makes too good a story for the news
media to let go.
5/24/2017
Machine Intelligence - web version
57
A Postscript on
Chess Machines
If Time Permits
5/24/2017
Machine Intelligence - web version
58
Did a Chess Machine beat the
human champion?
• The IBM Deep Blue team included two strong
chess players: Murray Campbell and, as a
consultant, Joel Benjamin, an international
grandmaster who had played Kasparov to a
draw in 1994.
• Members of the team describe their work as
using a computer to enhance the skills of a
human player. (It adds a few hundred points
to a person's chess rating.)
5/24/2017
Machine Intelligence - web version
59
Chess Playing Machines - 1
• Chess is a deterministic game, so a computer
could derive a winning solution analytically.
However the number of all possible positions
is so large (10120) that using even the fastest
available computer it will take billions of years
to consider all possible moves.
• Skilled players may look at 20 moves ahead
by pruning, i.e. ignoring non-promising
moves.
5/24/2017
Machine Intelligence - web version
60
Chess Playing Machines - 2
• Early efforts on computer chess were driven
by general AI methodology and focused on
imitating the human way of play.
• The research efforts were justified to the
funding agencies by the claim that playing
winning chess was just a special case of
general problem solving.
5/24/2017
Machine Intelligence - web version
61
Chess Playing Machines - 3
• Around 1980 Ken Thompson developed a
chess playing program called Belle based on
a minicomputer with a hardware attachment
used to generate moves very fast.
• Belle defeated all other computer programs
and became the world champion.
• The use of special chess knowledge and
special purpose hardware became the
preferred approach since then.
5/24/2017
Machine Intelligence - web version
62
Deep Blue
• A major focus of the effort was the
development of special purpose hardware (an
aspect discussed in a lecture given by one of
the designers at Stony Brook)
• A chess player (Murray Campbell )
contributed the evaluation functions of the
moves generated by the hardware.
5/24/2017
Machine Intelligence - web version
63
Deep Fritz, etc
• Deep Blue could examine 200 million moves
in a second (by using special purpose
hardware)
• Deep Fritz is software only and it can
examine only 2.5 million positions a second.
It tries to make up for its slower speed by
using pruning.
• In October 2002 it drew (4-4) with the current
world champion V. Kramnick.
5/24/2017
Machine Intelligence - web version
64