The Quest for Beneficial AI The Quest for Beneficial AI Eugénio Oliveira LIACC Faculty of Engineering, University of Porto CSCWD May 6, 2019
The Quest for Beneficial AI
The Quest for Beneficial AI
Eugénio OliveiraLIACC
Faculty of Engineering, University of Porto
CSCWD
May 6, 2019
The Quest for Beneficial AI
Towards a Beneficial AI
What is AI?
Eugénio Oliveira
Achievements
Further Research
Artificial General Intelligence ?
Promisses and Dangers
The Quest for Beneficial AI
Perceive the accessible Environment
Collect and interpret Data either structured or not
Infer Knowledge from Data
Apply Reasoning processes to Knowledge
Decide and take the best possible Action!
May learn and adapt its behaviour through the analysis of its action’s
outcomes in the environment
AI – Artificial Intelligence
Eugénio OliveiraEugénio Oliveira
The Quest for Beneficial AI
AI is not replacing designers.“If I considered the initials AI, I would have preferred Augmented Intelligence.”Former IBM CEO Ginni Rometty
Miklos PhilipsPrincipal User Experience and Interaction Designer London, UK
“Designers need not worry. Instead of being a threat, Augmented Intelligence will present a series of exciting opportunities... AI will make Designers smarter. We just need to wait a little longer…”
The Quest for Beneficial AI
Eugénio Oliveira
Statement 1: Our concern It is not about Robots. It is mostly about “Intelligent Software” developed in AI field.
Statement 2: Intelligence It is not only about the Mind. It is also about the huge sophistication of a multitude of processes going on in a BODY
the body is the support of the mind A mental structure would not be possible without a body structure, according to A. Damásio
The Quest for Beneficial AI
Eugénio Oliveira
Statement 3: Artificial intelligence IF used properly is useful to human kind.Intelligent robots work in conditions not suitable for human being:
Statement 4: AI as a Challenge:AI will ultimately emerge neither as a servant nor master technology, but as a human-AI cooperation
The Quest for Beneficial AI
Science Synergy will shape the future
Eugénio Oliveira
The Quest for Beneficial AI
What is really NEW with AI???
BIG DATA (Petabytes 10 15….)
DEEP LEARNING
Artificail Neural Networks based
Eugénio Oliveira
Summit supercomputer, at Oak Ridge National Lab. area: 2 tennis courts, >27,000 GPUs. AI/ Deep Learning for climate changes
HARDWARE: High Computational Power
The Quest for Beneficial AI
Symbolists
Connectionists
Evolutionaries
Statisticians (Bayesians)
Analogizers
The Master Algorithm: The ultimate Learning Machine that will remake our world
(P.Domingos, Basic Books, 2015)
Connectionists
The Quest for Beneficial AI
ei
= j=1
nwji *s
jsi =
11+exp(-ei )
or si = 12
*arctg(ei)
Transfer FunctionCombinationFunction
The Quest for Beneficial AI
The Quest for Beneficial AI
Multi Layer Perceptron in action
The Quest for Beneficial AI
AI achievements
Facebook decides which updates to show to the user and Twitter to decide which twits to show
Satellites to be intelligently managed
AI programs predict everything from our taste in music to our likelihood of committing a crime or our fitness for a job.
In 2006 R. Brooks opposed the idea that AI had failed and warned that AI would be around us every day in the next years.
Current AI-based systems allow:
Companies to record clients profile
The Quest for Beneficial AI
AI is BETTER …
Deep Blue and AlphaGo
(In October 2015, AlphaGo became the first Computer Go program to beat a professional human Go player without handicaps on a full-sized 19×19 board.)
The Quest for Beneficial AI
Deep Learning for speech recognition
Automatic TranslationEnglishMandarin and Mandarin English
AI achievements
The Quest for Beneficial AI
Google does page ranking for us
Computers perform Intelligent spam filtering
Amazon and Netflix Recommender Systems are in use
AI achievements
PANDORA records our preferred kind of music
from radio
BingTravel predicts future flights price
The Quest for Beneficial AI
AI achievements BDI Agents represent investors in the stock exchange
Skin steins can be automatically recognized as carcinogenic
AI in Predicting Bladder Cancer
INRIX helps to select the best route during traffic jams
The Quest for Beneficial AI
Machine Learning (ML) algorithms work together with a multitude of other algorithms in order to get the things done
Siemens Healthineers and IBM Watson Healthcombining the clinical expertise of Siemens and cognitive computing of IBM Watson Health to analyse critical healthcare data
IBM WATSON was tested for the diagnosis of 1000 cancersituations:
99% similar to the best human specialists.
Propose new unforeseen treatments in 34% of the situations by looking into available research articles
AI achievements
The Quest for Beneficial AI
“IBM Debater argued for governments subsidizing preschool. Professional human debater Harish Natarajan argued against”
returns ten most relevant articles from 4 million.Scanned 3,000 sentences in top ten articles
Assessed pro and con opinion polarity of candidateclaims. Constructed its speech with top claim predictions.
The Quest for Beneficial AI
GPT-2 may be used for:
Automatic text writing by an AI Assistant Automatic Machine Translation Better Voice Recognition
GPT-2 it is a language model trained with 8 million web pages dataset
OR for malicious purposes including:
Automatic Generation of “credible” FAKE NEWS Impersonate Humans online Automatic phishing
February 14, 2019
Better Language Models and Their Implications
OpenAI
Eugénio Oliveira
The Quest for Beneficial AI
“AI systems are already making problematicjudgements that are producing significant social, cultural, and economic impacts in people’s everyday lives.”
it is an iconic portrait of the indiscriminate horror
Why we urgently need to measure AI’s societal impactsBy Kate Crawford and Meredith Whittaker
The Quest for Beneficial AI
Super Inteligence? Artificial General Intelligence? Strong AI?
“Machine learning is remaking science, technology, business, politics, and war….” P.Domingos, TMA
“… and thanks to ML, which lets computers predict the future based on past experience the world will guess what you want before you even want it and will be ready for you” (P.Domingos
TEDxLA 2016).
Scaring?
Hyping AI
The Quest for Beneficial AI
AI: Achievements
Guohua Shen, Tomoyasu Horikawa, Kei Majima and Yukiyasu Kamitani
ATR Computational Neuroscience Laboratories, Kyoto University, Japan
Detect images/thaughts in theBrain ….
Retrieve Images through fMRI signals of Brain activity
Eugénio Oliveira
The Quest for Beneficial AI
through the “Master Algorithm", by mining a “vast amount of patient data and drug data combined with knowledge mined from biomedical literature is how we will cure the cancer".
If it exists, the Master Algorithm can derive all knowledge in the world—past, present, and future—from data. Inventing it would be one of the greatest advances in the history of science. P.D., TMA
“Big data is not the new oil; it's the new snake oil".
Hyping AI
The Quest for Beneficial AI
Dualism versus MonismCartesians are DualistsScience current trend is basically Monistic
Artificial General Intelligence raises the problem of a possible artificial Consciousness
I do not know. A real Digital Mind in the near future it is not probable.
But “Strong AI” can exist and raise many ethic problems anyway
The Mind is an emergent property of the Brain enabling humans to have a set of cognitive capabilities “The Digital
Mind”, Arlindo Oliveira
The Quest for Beneficial AI
To mimic a mind “in silico” Versus “in vivo”
Sentiments are different (but come after) Emotions (AD says)
Emotions can change the way Agents reason, memorize, act, decideWe may say Software agents act differently when they “feel like” happy or fearful or …
Watson won “Jeopardy“factoids. BUT…Was it happy after winning?
The Quest for Beneficial AI
Robot’s actions need: Reasoning to plan action and Reacting to the
environment
Problem: harmony between deliberation and reaction, preventing the robot from a possible schizoid character.
To represent inside our Robot's (agent) "mind", knowledge, believes, intentions and commitmentslinked with action. And also some kind of “Emotions”
“Emotion-like” states changed the way Agents reason, memorize, act, decide …
The Quest for Beneficial AI
ƐBDI LOGIC
The basic Emotional-BDI Logic system is characterised by a set of 10 Axioms:
DangerousThreat(Ψ,φ)≡FDES(φ)∧BEL(Ψ → AF(¬φ))∧BEL(AF Ψ)
StronglyUnpleasant(Ψ,φ) ≡ DES(φ) ∧ BEL(Ψ→ AF(¬ φ)) ∧ BEL(EF Ψ )
Threats: events in the environment which directly affect fundamental desires of the agent, putting at stake its self preservationDanger: a threat is dangerous when the agent believes that some condition falsifies a fundamental desire φ, and also believes that it will also be inevitably true in the future.
all propositional tautologies, time, action, belief, desire, intention, capabilities , resources, fear, fundamental desire
we define an ƐBDI-model with set of modal operators:
Op = {BEL,DES, INT, FDES, FEAR, ANX},
The Quest for Beneficial AI
Can a computer work like a Mind?
Download a mind to an artificial entity (robot, computer, computer network)? Reverse Engineering… ?
The right answer is not yet available !!!!
MRI, …, too much superficial.Resolution of 1mm3… 50-100K neurons, 102 to 103 synapses… We need interpretation at the Nano metric level
The Quest for Beneficial AI
Can a computer work like a Mind?
AlterEgo allows humans to converse in natural language with artificial intelligence assistants without any voice simply by vocalizing internally.A. Kapur, S. Kapur, and P. Maes, "AlterEgo: A Personalized Wearable Silent Speech Interface." 23rd IUI 2018, March 5, 2018.
The Quest for Beneficial AI
Can we be friends with a robot? John Danaher, Lecturer at National University of Ireland, Galway
Preconditions for such relationships:
Aristotelian friendship: relationship between two or more individuals through mutual sharing of values and interests, …
39
Mutuality condition: There must be mutual sharing of values and interests.
Honesty/authentic condition: The participants in the friendship must be honest with each other.
Equality condition: One party cannot think themselves superior to the other
Diversity condition: The participants must interact with one another in a varied and diverse set of circumstances
The Quest for Beneficial AI
Can you be friends with a robot?
Some say: Robots Cannot be Your Friends
40
Because: robots can NEVER satisfy the mutuality condition because robots can never have inner mental lives: they cannot truly share with us; The robot is too alien, too different, from us.
The Quest for Beneficial AI
Can you be friends with a robot?
Others say: All that matters is that they perform in such a way that we think they are satisfying those conditions.
41
We don’t have access to someone’s inner mental life; we can never know whether others really share our concerns, or whether they are representing themselves
all human-human friendships are founded on a performative/behaviourist satisfaction of the relevant conditions for friendship.
The Quest for Beneficial AI
“Human-like” AI?
Generalization
Learning
Knowledge
Transfer
The Role of sophisticated
BODY
Self-
counscienciousness
Common
sense
Eugénio Oliveira
DL networks
RL invent strategies
DeepMind IMPALA, transfers knowledge between tasks
The Quest for Beneficial AI
OBJECTIVE:
Maximize AI Benefits while Minimizing the Risks.
human-centric approach is needed!!
Attention to vulnerable groups: Child, aged people, patients, disabled, minorities.Attention to asymmetry in Power and Information between Employers and employees, Producers and Consumers …
Eugénio Oliveira
The Quest for Beneficial AI
May be the way to deal with the problem is not only
about technology.
May be it should be about Machine ethics.
How to make intelligent machines sharing our values?
What can we do?
Elon Musk at the World Government Summit:“ we may become irrelevant as artificial intelligence (AI) grows more prominent”, February 2017, Dubai
The Quest for Beneficial AI
Stuart Russell (U. California- Berkley) open letter ( FLI - Future of Life Institute ) to keep AI beneficial.
The Quest for Beneficial AI
It is naive to imagine that there will be just one set of ethical principles agreed by everybody.
Different Robot makers will produce different Ethic Codes
Need for a committee for AI and Robotics Ethics
• Certified and Ethically safe Software.
• Precise specifications since the beginning.
There will be, for sure, “Unfair-opportunistic Robots” in business and “Killer Robots”in the war.
Owners will argue that this only mimics the reality where you can find people that thrives because they are smarter or/and stronger.
The Quest for Beneficial AI
PAST EXPERIENCE (together with my students)
What can we do?
Dealing with critical situations (using MAS for Decision-Making):
to manage and coordinate actions: on a ship under attack When unexpected disruptions happen in an Airline OCC
it should be mandatory that the automated system specificationenforces the human in the loop. Someone who will be accountable forthe most important decisions in run-time.
The Quest for Beneficial AI
535353
Actors and goals diagram
Actors
AOSE helps on specifying Agents, hard/soft Goal, Actors, Roles andmandatory humaninteraction
The Quest for Beneficial AI
Appropriateness of Logic Programming-based reasoning to machine ethics with the capacity for “moral” decision making
“QUALM system mainly employs a combination of the Logic Programming features of abduction, updating, and counterfactuals.”
Research in progress
Probabilistic moral reasoning, to reason about actions, under uncertainty, that might have occurred,
This permits to capture a form of argumentation
The Quest for Beneficial AI
Abduction is a reasoning method whereby one chooses from available hypotheses those that best explain the observed evidence,
Research in progress
“Moral decisions are accomplished by satisfying a priori
integrity constraints, computing abductive stable models
from all possible abductive solutions, and then appropriately preferring amongst them (by means of rules)
The Quest for Beneficial AI
“knowledge base of the agent wrt this scenario:
divert(X) is an abducible, meaning ‘diverting object X’.
run_sidetrack(X) divert(X). hit(X, Y ) run_sidetrack(X), on_sidetrack(Y ).
save_from(X) sidetrack(straight), run_sidetrack(X).
save_from(X) sidetrack(loop), hit(X, Y ), heavy enough(Y ).
sidetrack(straight) foggy. sidetrack(loop)not foggy.
foggy. on_sidetrack(man). heavy_enough(man).
J. J. Alferes, L. M. Pereira, and T. Swift. Abduction in well-founded semantics and
generalized stable models via tabled dual programs. Theory and Practice of Logic
Programming, 4(4):383–428, 2004
Taking save_from(trolley) as the goal, the man hit by the trolley
is just a side-effect of achieving the goal, and thus divert(trolley)
is morally permissible
The Quest for Beneficial AI
AI Now Institute at New York University organized AI Now 2017 Symposium
- AI Now Report Ethical questions of bias and accountability will become even
more urgent … as AI systems capable of violent force against humans are developed … in military contexts.
There are difficult ethical issues involving how Lethal Autonomous Weapons Systems (LAWS) will detect threats or gestures of cooperation,
AI and robotics researchers should adopt ethical and legal standards that maintain human control and accountability over these systems.
The Quest for Beneficial AI
Starting a military AI arms race is a bad idea, and should be prevented by a ban on offensive autonomous weapons beyond meaningful human control.
Autonomous Weapons: an Open Letter from AI & Robotics Researchers http://tinyurl.com/awletter
The Quest for Beneficial AI
AI NOW Report
What can we do?
Near-term social and economic implications of AI addressing the following key issues:
Bias and inclusion: Bias that AI systems may perpetuatedue to biased training data ;
Labour Automation: impact on labour practices (employee hiring, firing and management);
Rights and Liberties: How can AI be used to support;
Ethics and Governance: How ethical codes could be developed in a time of political volatility
The Quest for Beneficial AI
Pew Research Center , Dec 10, 2018Artificial Intelligence and the Future of Humans
By Janna Anderson, Lee Rainie and Alex Luchsinger
The answers of the 979 respondents include:
63% who said most people will be better off
37% who said most people will not be better off
25 respondents who chose not to select either option
By 2030, do you think it is most likely that advancing AI andrelated technology systems will enhance human capacities andempower them?
The Quest for Beneficial AI
Q: As emerging algorithm-driven AI continues tospread, will people be better off than they are today?
Human agency : Individuals are experiencing a loss of control overtheir lives
Data abuse:(Our) Data use and surveillance in complex systems is designed forprofit or for exercising power
Job loss:The AI takeover of jobs will widen economic divides, leading tosocial upheaval despite the expected emergence of some new jobs
CONCERNS
The Quest for Beneficial AI
CONCERNS
Dependence lock-in:Reduction of individuals’ cognitive, social and survivalskills.
Mayhem:Autonomous weapons, cybercrime and weaponized information
The Quest for Beneficial AI
Alter economic and political systems to better help humans ‘racewith the robots’
SUGGESTED SOLUTIONS
Global good is No. 1:
Values-based system:
Prioritize people:
Improve human collaboration across borders and stakeholdergroups
Develop policies to assure AI will be directed at ‘humanness’ andcommon good
The Quest for Beneficial AI
Eugénio Oliveira
INDEPENDENT
HIGH-LEVEL EXPERT GROUP ON
ARTIFICIAL INTELLIGENCE SET UP BY THE EUROPEAN COMMISSION
ETHICS GUIDELINES
FOR TRUSTWORTHY AI
The Quest for Beneficial AI
Ethical Principles
By generating prosperity, value creation and wealthmaximization and sustainability;
By design, AI systems should protect the dignity, integrity,liberty, privacy, safety, and security of human beings
freedom from subordination to, or coercion by, AI systems.Human determination over themselves.
Eugénio Oliveira
The Principle of Beneficence: “DO GOOD”
The Principle of Non maleficence: “Do no Harm”
The Principle of Autonomy: “Preserve Human Agency”
The Quest for Beneficial AI
Developers and implementers need to ensure thatindividuals and minority groups maintain freedom frombias, stigmatisation and discrimination.
Technological transparency implies that AI systems be auditable,comprehensible and intelligible to human beings
Ethical Principles
The Principle of Justice: “Be Fair”
The Principle of Explicability: “Operate transparently”
The Quest for Beneficial AI
Recommendations
ART in ARTificial Intelligence: Accountability, Responsibility, Transparency
A: To whom should we address if an autonomous vehicle runs over a pedestrian?
R: System providers are responsible for the clarity of the decision-making.
T: System developers must guarantee the right specification, development and deployment good practices of the AI systems. To decide whether or not it is preferable to have the “Human in the Loop” and Ethical Software Constraints
The Quest for Beneficial AI
Say YES to some competences delegation! Obsolete humans ? Never!
Societies are able to survive and rebuild after economic/technological revolutions! But many concrete persons may be crushed in the process Our DUTY: do not allow that to happen
The Quest for Beneficial AI
THANK YOU!
Eugenio Oliveira([email protected])
AI is FASCINATING but …BE CAREFUL…