Possibly the most important lesson we have learned after 60 years of AI research is that what seemed to be very difficult to achieve, such as accurate medical diagnosis to playing chess at the level of a Grand Master, turned out to be relatively easy whereas what seemed easy, such as visual object recognition or deep language understanding, turned out to be extremely difficult. In my talk I will try to explain the reasons for this apparent contradiction by briefly reviewing the past and present of AI and projecting it into the near future.
Ramon Lopez de Mantaras is Research Professor of the Spanish National Research Council (CSIC) and Director of the Artificial Intelligence Research Institute of the CSIC. Technical Engineer EE (Electrical Engineering) from the Technical Engineering School of Mondragón (Spain) in 1973. Master of Sciences in Automatic Control from the University of Toulouse III (France) in 1974, Ph.D. in Physics from the University of Toulouse III (France), in 1977, with a thesis in Robotics (done at LAAS, CNRS). Master of Science in Engineering (ComputerScience) from the University of California at Berkeley (USA) in 1979. Ph.D. in Computer Science, from the Technical University of Catalonia, Barcelona (Spain) in 1981.
Past, Present and Future of AI: a Fascinating Journey - Ramon Lopez de Mantaras @ PAPIs Connect
1. Past, present, and future of AI:
A fascinating journey
Ramon Lopez de Mantaras
Artificial Intelligence Research Institute (IIIA)
CSIC
http://www.iiia.csic.es/~mantaras
2. Outline
- Turing on AI
From Turing to Dartmouth
- Two views on AI: Weak AI vs. Strong AI
-The road traveled
Achievements of (Weak) AI
-The (long) road ahead
From Integrated Systems to Strong AI
-Conclusions
3. Turing on AIIn 1948 Turing predicted that by the
end of the 20th century there would
be intelligent computers capable of
performing logical deductions, acquire
new knowledge inductively, by
experience and by evolution and
capable of communicating by means
of humanized interfaces. He also
speculated about a connection between
randomness and creative intelligence
by suggesting to add radium in to the
ACE in the hope that the random
decay of radiation would give its
inputs the desired unpredictability.
In his famous 1950 paper he also
speculated about the emulation of the
mind of a child and giving it an
appropriate education to obtain an
adult mind (mental development)
4. From Turing to Dartmouth
1948 Hixon Symposium on Cerebral Mechanisms in
Behavior in Caltech (McCulloch on NNs, von
Neumann, Lashley on limitations of behaviourism)
Session on Learning Machines at the 1955 Western
Computer Conference in L.A. (Clark & Farley on
Hebbian learning in NNs; Selfridge on image
classification; Newell on chess; Pitts on NNs)
1956 Summer Research Project on Artificial
Intelligence in Dartmouth College (McCarthy,
Minsky, Newell, Simon, Shaw, Selfridge,
Solomonoff, Rochester, Shannon, Samuel, Bernstein)
5. Two views on AI
-The view of the founding fathers:
The science and engineering of
replicating, even surpassing
(singularity?), human-level intelligence
in machines (“strong AI”)
-The view in the early 80’s (after the “AI
winter”):
The science and engineering
of designing machines with the
capability to perform tasks
that, when done by humans,
we agree that they require
intelligence (“weak AI”)
6. Strong versus Weak AI
The Strong AI case
Strong AI refers to AI that matches (or even exceeds)
general human-level intelligence (intelligent machines
will have mental states, consciousness, etc.)
Example: The robots from the movies (HAL
9000, Matrix, Terminator, I Robot, etc.)
The goal of human-level intelligence remains elusive but
has inspired and still inspires our work on AI even
though most efforts are on building weak AI (or “idiots
savants”)
7. Strong versus Weak AI
The Weak AI case (or the “idiots savants”)
Machines already exhibit specialized intelligences without
worrying about having mental states, consciousness, etc.
All current forms of AI are “weak AI”
We have achieved impressive results along the traveled
“weak AI” road
9. The road traveled
AI is everywhere (though most of the time is not visible!):
-Fuel injection systems in our cars designed using AI algorithms.
- Jet turbines are designed using genetic algorithms.
- 10.000 engineers carry out 2.600 maintenance works nightly on Hong Kong’s
subway, scheduled by an AI system
- There are a millions of AI-powered specialized robots in people’s homes and
robots running on the surface of Mars.
- Computer games (NPCs) use many AI techniques (including ML)
- Web search engines use AI techniques
- Automatic detection of credit card fraudulent transactions use ML algorithms
- Routing of cell phone calls is based on AI
- Detection of consumer habits is based on AI (ML)
- The world’s best chess players are computer programs
- Complex mathematical theorems have been proven by automatic theorem
provers (i.e. Robbins conjecture)
- An ML system revals passing patterns in soccer teams
- There are robots that play soccer
-There are AI systems composing beautiful music and systems
performing music expressively (among other artistic applications)
10. RoboCup: Learning to play cooperatively
R. Ros, R. Lopez de Mantaras, J.L. Arcos, M. Veloso; A Case-Based approach for Action Selection and Coodination
in Robot Soccer Gameplays, Artificial Intelligence Journal 173(9-10) (2009) 1014-1039.
doi:10.1016/j.artint.2009.02.004
12. The road traveled
We have achieved many of the things that the field’s founders used
as motivators, but not always in the way the “founding fathers” imagined:
-Very recently we have seen an impressive variety of application
achievements. Most of them based on the availability of very large sets of
data processed by very high performance computers, but NOT based on
emulating human’s mental processes:
-one of the world’s best Go players is a computer program
-self-driving cars have successfully run milions of miles (gathers 1 Gb/sec of data
to make predictions about its surroundings)
-there are high-performance speech recognition systems (SIRI, CORTANA,..)
-Watson outperformed the best “Jeopardy” players (and now… turns medic and
financial advisor)
-An ML system, trained on data from 133.000 patients from 4 Chicago’s
hospitals, can predict heart attacks in IC patients 4 hours before they happen
- …
13. The road traveled
In spite of all these great successes along specialized lines in each
of the areas of AI, we do not seem to be getting any closer to
“general AI” because of the following 3 problems:
1-We have given up the explainability of the AI systems (as well
as the cognitive plausability of AI models)
the “reasoning” made by today’s massive data-driven AI is
a massively complex statistical analysis of an immense
number of datapoints. We have traded the “why” for simply
the “what”
2-We have focused too much on the isolated components of AI
but not on the whole AI itself
We have wonderful bricks but, to build the house,
we need an architecture and the cement to tie the bricks
together (sensing, knowledge acquisition & representation,
reasoning, communication, action, planning, etc)
3- We have no idea of how to model and acquire common sense
knowledge
15. The road ahead: Integrated systems
Intelligence seems to emerge from a complex combination of many
specialized abilities, such as sensing, reasoning, learning, planning,
socializing, and communicating.
But not a mere juxtaposition of these abilities!
Rather, there is some set of deep interdependencies that tie these
elements together. For example:
-learning must result on knowledge that needs to be
represented so that reasoners, planners, etc can use it efficiently.
-perception requires reasoning and learning and
viceversa.
Most important challenge:
We need to think about how all the components of an artificial
intelligence should work together and how they need to be
connected (the architecture!). We need to focus on
comprehensive, totally integrated systems.
Integrated systems might be a necessary step towards strong
(human-level) AI (assuming this is a realistic goal!).
16. The road ahead
Example of Integrated System
Building a multipurpose, social, robot that can accumulate diverse
knowledge over long periods of time (continuous learning) and that can
use it effectively to decide what to do and how to do it.
Requirements
-A robot’s knowledge must be grounded in the physical world and capable
of learning by interacting with the world (“embodied cognition”)
-Because learning is prone to error, and the world is not deterministic,
reasoning with such learned knowledge must deal with uncertainty
-The representation languages must be expressive enough to represent the
complex connections between objects, places, actions, people, time, and
causation (understanding these requires common sense knowledge).
- Also requires deep natural language understanding (Watson does not
understand anything! neither does Google translator!) which depends on
common sense knowledge too!
- We need reliable computer vision systems capable of general object
recognition and deep scene understanding which again depends on
commons sense knowledge too!
17. Examples of the common sense knowledge that the
multipurpose social robot should have
• If a guest asks my waiter-robot for a glass of wine at a
party, and the robot sees the glass he is picked up is
cracked, or has a dead cockroach at the bottom, the robot
should not simply pour the wine into the glass and serve it.
• If a cat runs in front of my cleaning-robot while it is
cleaning my house, the robot should neither run it over nor
sweep it up nor put it away on a shelf.
But, unfortunately we are not quite there!
18. Big failures in scene understanding!
A red and white bus in front A man sitting in a bench with a dog
of a building
"a young boy is holding a baseball bat"
20. The road ahead. An alternative to the
common sense problem
The development and rapid deployment of
ubiquitous sensing and actuator devices makes it
possible to create AI systems robustly grounded in
direct experience with the world and learn
(including common sense knowledge) from
interacting with the world (i.e. work on
Developmental Robotics)
21. Developmental Robotics: Learning the musical
instrument and playing by imitation
(in collaboration with Imperial College)
A.Ribes, J. Cerquides, Y. Demiris, R. Lopez de Mantaras; Active Learning of Object and Body Affordances with Time
Constraints on a Humanoid Robot (in press) IEEE Transactions on Autonomous Mental Development
22.
23. The road ahead: Very ambitious predictions
-Robotic scientists that will serve as companions in discovery
by formulating hypothesis and pursuing their confirmation
(initial work on the ADAM and EVE systems by R. King et
al. "The Automation of Science". Science 324 (5923): 85–89)
-AI will play a central role in solving challenges in energy, the
environment, and in healthcare.
-A team of robots will beat the world’s human soccer
champion team. (H. Kitano)
-AI and other sciences (biology, material sciences,
nanotechnology, economics,…) will come together and will
have wide-ranging influences on our ideas about AI and on
the machines we will build.
24. Hydrogen muscle for silent robots
Copper and nickel-based metal hydride powder is compressed into peanut-sized pellets and
secured in a vessel. Hydrogen is pumped in to “charge” the pellets with the gas. A heater coil
surrounds the vessel. Heat breaks the weak chemical bonds and releases the stored hydrogen.
(Kim & Vanderhoff, Smart Mat. and Struct., 18, 2009 DOI: 10.1088/0964-1726/18/12/125014)
Inflatable rubber tube
surrounded by Kevlar
fibre braiding
25. Artificial cartilage
Chen, Briscoe, Armes, Klein; Lubrication at Physiological Pressures by Polyzwitterionic Brushes,
Science 323, 2009
Each molecular
group attracts 25
water molecules
Performs well in pressures up to 5 megapascals
60 nm backbone
26. Touch sensitive artificial skin
1977 2008
Capacitive copper contacts
A layer of silicone rubber acts as a spacer
between those contacts and an outer layer
of Lycra that carries a metal contact above
each copper contact. The whole constitutes
a pressure-sensing capacitor that can detect
a touch as light as 1 gram.
(Schmitz et al. IEEE Transactions on Robotics, 27(3).
2011)
Carbon, or metal, charged polymer coats the
fingers and palm. The transversal electrical
resistance varies as a function of the pressure.
Detected a touch greater than 20 grams. Applied
to tactile object recognition.
(López de Mántaras. PhD Thesis, Univ. Paul Sabatier.
1977)
28. Conclusions
-AI is a well stablished research discipline with demonstrated
successess and clear trajectories for its immediate future (but no
“singularity”: the brain is much too complex!).
-AI techniques are everywhere (although often are invisible): AI
Algorithms increasingly run our lives: They find books,
movies, jobs, and dates for us, manage our investments, and
discover new drugs.”
-Most exciting opportunities for research lie on the
interdisciplinary boundaries of AI with biology, linguistics,
economics, material sciences, etc. That will provide insights and
technologies towards building large-scale integrated systems.
-AI is mature enough to undertake again the research on cognitive
architectures and integrated systems (perhaps leading towards
the goals of more general and strong AI?) and not only working
on massive data-driven AI (which will NOT lead us towards
general and strong AI).
29. BUT…
…progress will be slow because:
The problem of common sense knowledge is
much too hard
The field is much too dominated by massive data-
driven AI
…and because AI suffers from fragmentation
(separate conferences and over-specialized college
curricula)
However the progress in weak AI will continue to
be formidable
30. Final thoughts
No matter how sophisticated will future Artificial
Intelligences be they will necessarily be different to human
intelligences because:
THE BODY SHAPES THE WAY WE THINK
These artificial intelligences will be alien to human needs and
therefore we should put limits on the developments of AI,
particularly in fully autonomous (and therefore
uncontrolable!) systems
In any case…”KEEP CALM AND FORGET ABOUT
THE SINGULARITY”