AI: Paradigm Shifts

1
AI Paradigm Shifts

• AI research trends continue to shift
• Moving AI from a stand-alone component, to a
  component within other software systems
• consider the original goal was to build
  intelligence
• later, the goal became problem solving systems
• Now the goal is autonomous (agents) or
  semi-autonomous (robots)
• software systems or systems that work with humans
  (data mining, decision support tools)
• Machine learning has similarly shifted
• Originally the concept was vague with no ideas of
  how to approach it
• early symbolic approaches dealt with acquiring
  knowledge or building on top of already present
  knowledge
• neural networks focused on training how to solve
  a given task
• Today, we often look at learning as improving
  performance through training

2
Other Paradigm Shifts

• Early AI was almost solely symbolic based
• Early neural network research of the 1960s made
  no impact
• In the 1980s, there was a shift away from
  symbolic to connectionism
• But that shift was somewhat short-lived as neural
  network limitations demonstrated
• Today, we see all kinds of approaches
• Symbolic knowledge-based
• possibly using fuzzy logic
• Symbolic ontologies
• Symbolic probabilistic through networks
  (Bayesian, HMM)
• Neural network
• Genetic algorithms

3
AI Approaches in the Future

• Obviously, we cant predict now what approaches
  will be invented in the next 5-10 years or how
  these new approaches will impact or replace
  current approaches
• However, the following approaches are finding
  uses today and so should continue to be used
• data mining on structured data
• machine learning approaches Bayesian and neural
  network, support vector machines
• case based reasoning for planning, model-based
  reasoning systems
• rules will continue to lie at the heart of most
  approaches (except neural networks)
• mathematical modeling of various types will
  continue, particularly in vision and other
  perceptual areas
• Interconnected (networks) software agents
• AI as part of productivity software/tools

4
AI Research For the Short Term

• Reinforcement learning
• Applied to robotics
• Semantic web
• Semi-annotated web pages
• Development of more intelligent agents
• Speech recognition
• Improvement in areas of accuracy, larger
  vocabulary, and speed (reducing amount of search)
• Natural language understanding
• Tying symbolic rule-based approaches with
  probabilistic approaches, especially for semantic
  understanding, discourse and pragmatic analysis

5
Continued

• Social networks
• Modeling and reasoning about the dynamics of
  social networks and communities including email
  analysis and web site analysis
• Multi-agent coordination
• How will multiple agents communicate, plan and
  reason together to solve problems such as
  disaster recovery and system monitoring (e.g.,
  life support on a space station, power plant
  operations)
• Bioinformatics algorithms
• While many bioinformatics algorithms do not use
  AI, there is always room for more robust search
  algorithms

6
Some Predictions?

• Next 5-10 years
• Work continues on semantic web,
  robotics/autonomous vehicles, NLP, SR, Vision
• Within 10 years
• part of the web is annotated for intelligent
  agent usage
• modest intelligent agents are added to a lot of
  applications software
• robotic caretakers reach fruition (but are too
  expensive for most)
• SR reaches a sufficient level so that continuous
  speech in specific domains is solved
• NLP in specific domains is solved
• Reliable autonomous vehicles used in specialized
  cases (e.g., military)

7
And Beyond

• Within 20 years
• robotic healthcare made regularly available
• vision problem largely solved
• autonomous vehicles available
• intelligent agents part of most software
• cognitive prosthetics
• semantic web makes up a majority of web pages
• computers regularly pass the Turing Test
• Within 50 years
• nano-technology combines with agent technology,
  people have intelligent machines running through
  their bodies!
• humans are augmented with computer memory and
  processors
• computers are inventing/creating useful artifacts
  and making decisions
• Within 100 years (?)
• true (strong) AI

8
Other Predictions

• Want to place a bet? These bets are available
  from www.longbets.org/bets
• By 2020, wearable devices will be available that
  will use speech recognition to monitor and index
  conversations and can be used as supplemental
  memories by Greg Webster (??)
• By 2025, at least half of US citizens will have
  some form of technology embedded in their bodies
  for ID/tracking Douglas Hewes (CEO Business
  Technologies)
• By 2029, no computer will have passed the Turing
  test by Ray Kurzweil (a well known entrepreneur
  and technologist)
• By 2030, commercial passenger planes will fly
  pilotless by Eric Schmidt (CEO Google)
• By 2050, no machine intelligence will be
  self-aware by Nova Spivack (CEO of Lucid
  Ventures)
• By 2108, a sentient AI will exist as a
  corporation providing services as well as making
  its own financial and strategic decisions by
  Jane Walter (??)

9
Wearable AI

• Wearable computer hardware is becoming more
  prevalent in society
• we want to enhance the hardware with software
  that can supply a variety of AI-like services
• The approach is called humanistic intelligence
  (HI)

• HI includes the human in the processing such that
  the human is not the instigator of the process
  but the beneficiary of the results of the process

10
HI Embodies 3 Operational Modes

• Constancy the HI device is always operational
  (no sleep mode) with information always being
  projected (unlike say a wrist watch where you
  have to look at it)
• Augmentation the HI augments the humans
  performance by doing tasks by itself and
  presenting the results to the human
• Mediation the HI encapsulates the human, that
  is, the human becomes part of the apparatus for
  instance by wearing special purpose glasses or
  headphones (but the HI does not enclose the
  human)
• These systems should be unmonopolizing,
  unrestrictive, observable, controllable,
  attentive, communicative

11
HI Applications

• Filtering out unwanted information and alerting
• specialized glasses that hide advertisements or
  replace the content with meaningful information
  (e.g., billboards replaced with news)
• blocking unwanted sounds such as loud noises with
  white noise
• alerting a driver of an approaching siren
• providing GPS directions on your glasses
• Recording perceptions
• if we can record a persons perceptions, we might
  be able to play them back for other people
  record a live performance
• other examples include recording people
  performing an activity so that it can be repeated
  (by others)
• record hand motions while the user plays piano
• record foot motions while the user dances to
  capture choreography

12
Continued

• Military applications
• aiming missiles or making menu selections in an
  airplane so that the pilot doesnt have to move
  his hands from the controls some of this
  technology already exists
• reconnaissance by tracking soldiers in the field,
  seeing what they are seeing
• Minimizing distractions
• using on-board computing to determine what a
  distraction might be to you and to prevent it
  from arising or blocking it out
• Helping the disabled
• HI hearing aids, HI glasses for filtering,
  internal HI for medication delivery, reminding
  and monitoring systems for the elderly

13
Beyond AI Wearables

• As the figure below shows, these devices may be
  more intimately wound with the human body
• We are currently attaching ID/GPS mechanisms to
  children and animals
• Machine-based tattoos are currently being
  researched

• What about underneath the skin?
• Nano-technology
• Hardware inside the human body (artificial
  hearts, prosthetic device interfaces, etc)

14
AI in Space/NASA

• Planning/scheduling
• Manned mission planning, conflict resolution for
  multiple missions
• Multi-agent planning, distributed/shared
  scheduling, adaptive planning
• Rover path planning
• Telescope scheduling for observations
• Deliberation vs. reactive control/planning
• Plan recovery (failure handling)
• Life support monitoring and control for safety
• Simulation of life support systems
• On-board diagnosis and repair
• Science
• Weather forecasting and warning, disaster
  assessment
• Feature detection from autonomous probes
• Other forms of visual recognition and discovery

15
Smart Environments

• Sometimes referred to as smart rooms
• Components collection of computer(s), sensors,
  networks, AI and other software, actuators (or
  robots) to control devices
• Goal the environment can modify itself based on
  user preferences or goals and safety concerns
• smart building might monitor for break-ins, fire,
  flood, alert people to problems, control traffic
  (e.g. elevators)
• smart house might alter A/C, adjust lighting,
  volume, perform household chores
  (starting/stopping the oven, turn on the
  dishwasher), determine when (or if) to run the
  sprinkler system for the lawn
• smart restaurant might seat people automatically,
  have robot waiters, automatically order food
  stock as items are getting low (but not actually
  cook anything!)

16
Smart Windows

• One of the more imminent forms of smart
  environments is the smart window
• To help control indoor environments as an
  energy-saving device
• The window contains several sheets of optical
  film, each of which is controlled by a roller
  that can roll the film up or down
• there are six microcontrollers
• presumably the system works by fuzzy logic
  although I could not find such details on how the
  controllers made decisions
• An optical sensor allows the window to identify
  the current situation (too much light, too much
  heat, not enough heat) and respond by creating
  the needed level of translucency by sliding films
  up or down

17
Smart Room
18
Automated Highways

• Features
• Provide guidance information for cooperative
  (autonomous) vehicles
• Monitor and detect non-cooperative vehicles and
  obstacles
• Plan optimum traffic flow
• Architecture
• Network of short-range hi-resolution radar
  sensors on elevated poles
• Additional equipment in vehicles (transponders
  for instance for location and identification)
• Sensors on the road for road conditions and on
  the vehicles for traction information
• Sensors for other obstacles (e.g., animals)
• Computer network
• Roadway blocked off from sidewalk and pedestrian
  traffic

19
Evolution of AVs/Highways
20
Smart Highway
21
Smart City Block
22
Creating Human-level Intelligence

• This was our original goal
• Is it still the goal of AI?
• Should this be the primary goal of AI?
• What approaches are taking us in that direction?
• Cyc?
• Cog and other efforts from Brooks?
• Semantic web and intelligent agents?
• What do we need to improve this pursuit?
• Study the brain? Study the mind?
• Study symbolic approaches? Subsymbolic
  approaches?
• Machine learning?
• In spite of such pursuits, most AI is looking at
  smaller scale problems and solutions
• And in many cases, we now are willing to embrace
  helper programs that work with human users

23
Social Concerns Unemployment

• According to economics experts, computer
  automation has created as many jobs as it has
  replaced
• Automation has shifted the job skills from blue
  collar to white collar, thus many blue collar
  jobs have been eliminated (assembly line
  personnel, letter sorters, etc)
• What about AI?
• Does AI create as many jobs as it makes obsolete?
• probably not, AI certainly requires programmers,
  knowledge engineers, etc, but once the system is
  created, there is no new job creation
• Just what types of jobs might become obsolete
  because of AI?
• secretarial positions because of intelligent
  agents?
• experts (e.g., doctors, lawyers) because of
  expert systems?
• teachers because of tutorial systems?
• management because of decision support systems?
• security (including police), armed forces,
  intelligence community?

24
Social Concerns Liability

• Who is to blame when an AI system goes wrong?
• Imagine these scenarios
• autonomous vehicle causes multi-car pile-up on
  the highway
• Japanese subway car does not stop correctly
  causing injuries
• expert medical system offers wrong diagnosis
• machine translation program incorrectly
  translating statements between diplomats leading
  to conflict or sanctions
• We cannot place blame on the AI system itself
• According to law, liability in the case of an AI
  system can be placed on all involved
• the user(s) for not using it correctly
• the programmers/knowledge engineers
• the people who supplied the knowledge (experts,
  data analysts, etc)
• management and researchers involved
• AI systems will probably require more thorough
  testing than normal software systems
• at what point in the software process should we
  begin to trust the AI system?

25
Case Study Therac-25

• Medical accelerator system to create high energy
  electron beams
• used to destroy tumors, can convert the beam to
  x-ray photons for radiation treatments
• Therac-25 is both hardware and software
• earlier versions, Therac-6 and Therac-20, were
  primarily the hardware, with minimal software
  support

• Therac-6 and -20 were produced by two companies,
  but Therac-25 was produced only be one of the two
  companies (AECL) , borrowing software routines
  from Therac-6 (and unknown to the quality
  assurance manager, from Therac-20)
• 11 units sold (5 in US, 6 in Canada) in the early
  to mid 80s, during this time, 6 people were
  injured (several died) from radiation overdoses

26
The 6 Reported Accidents

• 1985 woman undergoing lumpectomy receives
  15,000-20,000 rads eventually she loses her
  breast due to over exposure to radiation, also
  loses ability to use arm and shoulder
• treatment printout facility of Therac-25 was not
  operating during this session and therefore AECL
  cannot recreate the accident
• 1985 patient treated for carcinoma of cervix
  user interface error causes overexposure of
  1317,000 rads, patient dies in 4 months of
  extremely virulent cancer, had she survived total
  hip replacement surgery would have been needed
• 1985 treatment for erythema on right hip
  results in burning on hip, patient still alive
  with minor disability and scarring

27
Continued

• 1986 patient receives overdoes caused by
  software error, 16,500-25,000 rads, dies within 5
  months
• 1986 same facility error, patient receives
  25,000 rads dies within 3 weeks
• 1987 AECL had fixed all of the previously
  problems, new error of hardware coupled with user
  interface and operator error results in a
  patient, who was supposed to get 86 rads being
  given 8-10,000 rads, patient dies 3 months later
• note Therac-20 had hardware problems which
  would have resulted in the same errors from
  patients 4 and 5 above, but because the safety
  interlocks were in hardware, the error never
  arose during treatment to harm a patient

28
Causes of Therac-25 Accidents

• Therac-20 used hardware interlocks for
  controlling hardware settings and ensuring safe
  settings before beam was emitted
• User interface was buggy
• Instruction manual omitted malfunction code
  descriptions so that users would not know why a
  particular shut down had occurred
• Hardware/software mismatch led to errors with
  turntable alignment
• Software testing produced a software fault tree
  which seemed to have made up likelihoods for
  given errors (there was no justification for the
  values given)

29
Continued

• In addition, the company was slow to respond to
  injuries, and often reported we cannot recreate
  the error, they also failed to report injuries
  to other users until forced to by the FDA
• Investigators found that the company had less
  than acceptable software engineering practices
• Lack of useful user feedback from the Therac-25
  system when it would shut down, failure reporting
  mechanism off-line during one of the accidents

30
Safety Needs in Critical Systems

• It is becoming more and more important to apply
  proper software engineering methodologies to AI
  to ensure correctness
• Especially true in critical systems (Therac-25,
  International Space Station), real-time systems
  (autonomous vehicles, subway system)
• Some suggestions
• Increase the usage of formal specification
  languages (e.g., Z, VDM, Larch)
• Add hazard analysis to requirements analysis
• Formal verification should be coupled with formal
  specification
• statistical testing, code/document inspection,
  automated theorem provers
• Develop techniques for software development that
  encapsulate safety
• formal specifications for component retrieval
  when using previously written classes to limit
  the search for useful/usable components
• reasoning on externally visible system behavior,
  reasoning about system failures (this is
  currently being researched to be applied to life
  support systems on the International Space
  Station)

31
Social Concerns Explanation

• An early complaint of AI systems was their
  inability to explain their conclusions
• Symbolic approaches (including fuzzy logic, rule
  based systems, case based reasoning, and others)
  permit the generation of explanations
• depending on the approach, the explanation might
  be easy or difficult to generate
• chains of logic are easy to capture and display
• Neural network approaches have no capacity to
  explain
• in fact, we have no idea what internal nodes
  represent
• Bayesian /HMM approaches are limited to
• probabilistic results (show probabilities to
  justify answer)
• paths through an HMM

32
Continued

• As AI researchers have moved on to more
  mathematical approaches, they have lost the
  ability (or given up on the ability) to have the
  AI system explain itself
• How important will it be for our AI system to
  explain itself?
• is it important for speech recognition?
• is it important for an intelligent agent?
• here, the answer is probably yes, if the agent is
  performing a task for a person, the person may
  want to ask why did you choose that?
• is it important for a diagnostic system?
• extremely important
• is it important for an autonomous vehicle?
• possibly only for debugging purposes

33
Social Concerns AI and Warfare

• What are the ethics of fighting a war without
  risking our lives?
• Consider that we can bomb from a distance without
  risk to troops since this lessens our risk,
  does it somehow increase our decision to go to
  war?
• How would AI impact warfare?
• mobile robots instead of troops on the
  battlefield
• predator drone aircraft for surveillance and
  bombing
• smart weapons
• better intelligence gathering
• While these applications of AI give us an
  advantage, might they also influence our decision
  to go to war more easily?
• On the other hand, can we trust our fighting to
  AI systems?
• Could they kill innocent bystanders?
• Should we trust an AI systems intelligence
  report?

34
Social Concern Security

• In a similar vein, we are attempting to use AI
  more and more in the intelligence community
• Assist with surveillance
• Assist with data interpretation
• Assist with planning
• Will the public back AI-enhanced security
  approaches?
• What happens if we come to rely on such
  approaches?
• Are they robust enough?
• Given the sheer amount of data that we must
  process for intelligence, AI approaches makes
  fiscal sense
• How do we ensure that we do not have gaps in what
  such systems analyze
• How do we ensure accuracy of AI-based
  conclusions?
• In some ways, we might think of AI in security as
  a critical system, and AI in disaster planning as
  a real time system

35
Social Concerns Privacy

• This is primarily a result of data mining
• We know there is a lot of data out there about us
  as individuals
• what is the threat of data mining to our privacy?
• will companies misuse the personal information
  that they might acquire?
• We might extend our concern to include
  surveillance why should AI be limited to
  surveillance on (hypothetical) enemies?
• Speech recognition might be used to transcribe
  all telephone conversations
• NLU might be used to intercept all emails and
  determine whether the content of a message is
  worth investigating
• We are also seeing greater security mechanisms
  implemented at areas of national interests
  (airports, train stations, malls, sports arenas,
  monuments, etc) cameras for instance
• previously it was thought that people would not
  be hired to watch everyone, but computers could

36
What If Strong AI Becomes a Reality?

• Machines to do our work for us leaves us with
• more leisure time
• the ability to focus on educational pursuits,
  research, art
• computers could teach our young (is this good or
  bad?)
• computers could be in charge of transportation
  thus reducing accidents, and possibly even saving
  us on fuel
• computers may even be able to discover and create
  for us
• cures to diseases, development of new power
  sources, better computers
• On the negative side, this could also lead us
  toward
• debauchery (with leisure time we might degrade to
  decadence)
• consider ancient Romans had plenty of free time
  because of slavery
• unemployment which itself could lead to economic
  disaster
• if computers can manufacture for us anything we
  want, this can also lead to economic problems
• We might become complacent and lazy and therefore
  not continue to do research or development

37
AI The Moral Dilemma

• Researchers (scientists) have often faced the
  ethical dilemmas inherent with the product of
  their work
• Assembly line
• Positive outcomes increased production and led
  to economic boons
• Negative outcomes increased unemployment,
  dehumanized many processes, and led to increased
  pollution
• Atomic research
• Positive outcomes ended world war II and
  provided nuclear power,
• Negative outcomes led to the cold war and the
  constant threat of nuclear war, creates nuclear
  waste, and now we worry about WMDs
• Many researchers refused to go along with the US
  governments quest to research atomic power once
  they realized that the government wanted it for
  atomic bombs
• They feared what might come of using the bombs
• But did they have the foresight to see what other
  problems would arise (e.g., nuclear waste) or the
  side effect benefits (eventually, the arms race
  caused the collapse of the Soviet Union because
  of expense)
• What side effects might AI surprise us with?

38
Long-term Technological Advances

• If we extrapolate prior growth of technology, we
  might anticipate
• enormous bandwidth (terabit per second),
  secondary storage (petabyte) and memory
  capacities (terabyte) by 2030
• in essence, we could record all of our
  experiences electronically for our entirely
  lifetime and store them on computer, we can also
  download any experience across a network quickly
• Where might this lead us?
• Teleportation combining network capabilities,
  virtual reality and AI
• Time travel being able to record our
  experiences, thoughts and personalities, in a
  form of agent representative, so that future
  generations can communicate with us combining
  machine learning, agents, NLU
• Immortality the next step is to then upload
  these representatives into robotic bodies, while
  these will not be us, our personalities can live
  on, virtually forever

39
Ethical Stance of Creating True AI

• Today we use computers as tools
• software is just part of the tool
• AI is software
• will we use it as a tool?
• Does this make us slave masters?
• ethically, should we create slaves?
• if, at some point, we create strong AI, do we set
  it free?
• what rights might an AI have?
• would you permit your computer to go on strike?
• would you care if your computer collects data on
  you and trades it for software or data from
  another computer?
• can we ask our AI programs to create better AI
  programs and thus replace themselves with better
  versions?
• What are the ethics of copying AI?
• we will presumably be able to mass produce the AI
  software and distribute it, which amounts
  essentially to cloning
• humans are mostly against human cloning, what
  about machine cloning?

40
End of the World Scenario?

• When most people think of AI, they think of
• AI run amok
• Terminator, Matrix, etc (anyone remember
  Colossus The Forbin Project?)
• Would an AI system with a will of its own
  (whether this is self-awareness or just
  goal-oriented) want to take over mankind or kill
  us all?
• how plausible are these scenarios?
• It might be equally likely that an AI that has a
  will of its own would just refuse to work for us
• might AI decide that our problems/questions are
  not worthy of its time?
• might AI decide to work on its own problems?
• Can we control AI to avoid these problems?
• Asimovs 3 laws of robotics are fiction, can we
  make them reality?
• How do we motivate an AI? How do we reward it?