Artificial Intelligence Introduction PDF

Artificial Intelligence Introduction  75% Attendance Late entrance Course materials Elaine Rich and Kevin Knight, Artificial Intelligence, Tata McGraw Hill. Artificial Intelligence: a modern approach, Peter Norvig and Stuart J. Russell  Histo AI What ry of T...

Artificial Intelligence Introduction  75% Attendance Late entrance Course materials Elaine Rich and Kevin Knight, Artificial Intelligence, Tata McGraw Hill. Artificial Intelligence: a modern approach, Peter Norvig and Stuart J. Russell  Histo AI What ry of Toda ? is AI? AI y AI Ethics & Safety What is AI? ASIMO (Advanced Step in Innovative Mobility) is a humanoid robot created by Honda in 2000 What is Intelligence? Definition from Merriam-Webster  What is AI? The science of making machines that: Think like Think people rationally Act like people Act rationally  Rational Decisions We’ll use the term rational in a very specific, technical way:  Rational: maximally achieving pre-defined goals  Rationality only concerns what decisions are made (not the thought process behind them)  Goals are expressed in terms of the utility of outcomes  Being rational means maximizing your expected utility The Goal of AI “Have machines solve problems that are challenging for humans.” We call such a machine an intelligent agent. Narrow AI Artificial general intelligence An intelligent A hypothetical agents can solve a (AGI)intelligent agent specific problem. which can understand or learn any intellectual task that human beings can. [Wikipedia entry on AGI] How can we achieve this? Create an agent that can Think Think Act Act like a like a rationall rationall human? human? y? y?  Think Think Act Act like a like a rationall rationall human? human? y? y? The brain as an How to understand information cognition as a AI consciousness processing computational machine. process? Requires scientific Introspection: try What does it theories of how to think about mean that a the brain works. how we think. machine is Predict the conscient/sentient behavior of ? Note: The brain human subjects. How can we tell? does not work like Image the brain, artificial neural examine networks from neurological data (What do we do?) ML! Cognitive Sciences  Think Think Act Act like a like a rationall rationall human? human? y? y? Alan Turing rejects the question “Can machines think?” The Turing Test tries to define what acting like a human means Alan Turing (1950) "Computing machinery and intelligence“ What capabilities would a computer need to have to pass the Turing Test? These are still the core AI areas. Natural language processing Knowledge representation Automated reasoning Machine learning Turing predicted that by the year 2000, machines would be able to fool 30% of human judges for five minutes. Turing Test: Criticism What are some potential problems with the Turing Test? Chinese Room Some human behavior is not intelligent. Argument Some intelligent behavior may not be human. Human observers may be easy to fool. A lot depends on expectations. Anthropomorphic fallacy: humans tend to humanize things. Imitate intelligence without intelligence. E.g., the early chatbots ELIZA (1964) simulates a conversation using pattern matching. Thought experiment by John Searle (1980): Imitate intelligence using Is passing the Turing test a good rules. scientific goal? Engineering perspective: Imitating a human is not a good way to solve practical problems. We can create useful intelligent agents without trying to imitate humans.  Think Think Act Act like a like a rationall rationall human? human? y? y? Thinking Rationality: Draw sensible conclusions from facts, logic and data. Logic: A chain of argument that always yield correct conclusions. E.g., “Socrates is a man; all men are mortal; therefore, Socrates is mortal.” Logic-based approach to AI: Describe problem in formal logic notation and apply general deduction procedures to solve it. Issues: Describing real-world problems and knowledge using logic notation is hard. Computational complexity of finding the solution. Much intelligent or “rational” behavior in an uncertain world cannot be defined by simple logic rules. What about the logical implication Should it rather be  Think Think Act Act like a like a rationall rationall human? human? y? y? Acting rational means to try to achieve the “best” outcome. Best means that we need to do optimization. The desirability of outcomes can be measured by the economic concept of utility. If there is uncertainty about achieving outcomes, then we need to maximizing the expected utility. Optimization has several advantages: Generality: optimization is not limited to rules. Practicality: can be adapted to many real-world problems. Well established: solvers, simulation and experimentation. Avoids philosophy and psychology in favor of a clearly defined objective. Bounded rationality: In practice, expected utility optimization is subject to the agent’s knowledge and computational constraints.  What type of AI do we cover in this course? Create a narrow AI agent that can think like think act act like a a rationall rational human? human? y? ly. That is, use machines that act in a way to solve a specific hard problem that traditionally would have been thought to require human intelligence. What are the Components of an Intelligent Agent? Intelligent agents Developer need to Communicate with the environment. Represent knowledge, reason and plan to achieve a desired outcome. Agent interacting with the environment [Artificial Intelligence: A Modern Approach, Editions Optional 1-3] Learn to improve performance. Example: Self-Driving Car Percept People crossing the street : Action: Stop the car Example: LLM Percept prompt : Action: next most likely word More words are created word-by- word. Machine Learning vs. Artificial Intelligence AI Designing an intelligent agent Vision Sensors M Motion NLP L Learning from examples Knowledge instead of being programmed representatio Supervised learning n Unsupervised ML Planning Deep learning Goals RL Learnin g ML ML Agent interacting with the environment [Artificial Intelligence: A Modern Approach, Editions 1-3] The History of AI  1974-1980 First AI Winter 1987-1993 Second AI Winter 1989: Universal approximation theorem for neural networks. 2010 2022 2017 2015 Transformer Generative AI architecture models: and large DALL-E language ChatGPT, models Gemini Deep Learning Revolution LLMs … (learning layered artificial neural networks) starts fueled by NVIDIA GPUs. Now Google enables leaps in image processing and speech recognition. Source: https://qbi.uq.edu.au/brain/intelligent-machines/history-artificial-intelligence + add What accounts for recent successes in AI? Faster computers and specialized hardware (GPUs). Lots of data (the Internet, text, sensors) and storage (cloud) Dominance of machine learning. New optimization methods (deep learning). “Moravec’s Paradox” Hans Moravec (1988): “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.” A teenager can learn how to drive in a few hours with very little input, but we still have no truly self-driving car. https://www.newsweek.com/googles-new-two-legged-r obot-future-warfare-429831 The AI Effect: AI gets no respect? As soon as a machine gets good at performing some task, the task is no longer considered to require much intelligence Calculating ability used to be prized – not anymore. Chess was thought to require high intelligence – now computers play at a super-human level. Learning once thought uniquely human - now machine learning is a well-developed discipline. Art? “Even a monkey can do this!” AI Today Vision and Image Processing OCR: read license plates, handwriting recognition (e.g., mail sorting). Face detection: now standard for DALL-E smart phone cameras. Vehicle safety systems Visual search Image generation All these technologies operate now at superhuman performance. Natural Language Processing Text-to-speech Speech-to-text to detect voice commands Machine translation Text generation (Q/A systems) using Large Language Models These technologies operate now with close to or even superhuman performance. Humans use language to reason. Does that mean AI that can create good language can reason? Language understanding is still elusive! Robotics Mars rovers Autonomous vehicles DARPA Grand C hallenge Google self- driving cars Autonomous helico pters and drones Robot soccer RoboCup Personal robotics Humanoid robots Robotic pets Personal assistants? Question Answering: IBM Watson Listens to spoken http://www.research.ibm.com/deepq a/ language. NY Times article Speaks. Trivia demo Finds questions YouTube video to factual answers. IBM Watson wins on Jeopardy (February 2011) Math, Games and Puzzles 1996: A computer program written by researchers at Argonne National Laboratory proved a mathematical conjecture (Robbins conjecture) unsolved for decades NY Times story: “[The proof] would have been called creative if a human had thought of it” 1996/97: IBM’s Deep Blue defeated the reigning world chess champion Garry Kasparov in 1997 1996: Kasparov Beats Deep Blue “I could feel --- I could smell --- a new kind of intelligence across the table.” 1997: Deep Blue Beats Kasparov “Deep Blue hasn't proven anything.” 2007: Checkers was “solved” --- a computer system that never loses was developed. Science article 2017+: AlphaZero learns chess, shogi and go by playing itself. Science article 2019: MuZero learns to play Atari computer games. AI exhibits superhuman performance on almost all games. AI Ethics & Safety A new Frontier for Fairness and Freedom AIMA Chapter 27 Commonly-Cited Safety and Ethics Principles Use of AI by Ensure safety companies Limit harmful uses of AI and Establish accountability: Liability? organization Avoid concentration of power: Winner- s takes-All Uphold human rights and values Ensure fairness: Equal opportunity/equal impact. Reflect diversity/inclusion Protect Provide transparency: Explanations to individuals build trust Respect privacy: Surveillance? Contemplate implications for employment: Income and purpose. Governance Acknowledge legal/policy implications Next, we look at the implementation of these principles in different countries.  European 20 Union 6 1 Has regulations since 2016 included in the General Data Protection Regulation Art. 22 GDPR – Automat ed individual decision-m (GDPR) aking, including California’s CCPA was not modeled after the GDPR  20 9 1 European Union Study https://www.europarl.europa.eu/thinktank/en/document.html?reference=EPRS_STU(2019) 624262  20 1 2 Source: https://ai.google/static/documents/recommendations-for-regulating-ai.pdf  20 US White House Executive Order 3 2 14110 Some important points: Artificial Intelligence must be safe and secure. Promoting responsible innovation, competition, and collaboratio n Americans’ privacy, civil liberties and labor rights must be protected. Algorithmic Bias “Algorithmic bias describes systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others.” [Wikipedia] Pre-existing Technical Emergent b bias Social and institutional norms influence design bias Limitations program or of a ias Use and reliance on algorithms across new and training data computational power. or unanticipated choices. contexts. Example: Evaluate job Example: instead of a Example: Use of an applicants for a job random sample, the algorithm for an which is historically program uses the first unanticipated almost exclusively held n data points. application that would by males. require retraining. AI Safety “Prevent accidents, misuse, or other harmful consequences of AI.” AI Testing Monitoring Adversarial AI robustness How should this be ensured? Corporate self-regulation Government action Credit: Terminator 3: Rise of the Machines. Warner Br AI Safety Intelligent Agents are “optimizers!” Goal/reward alignment: How do we specify a robust objective function? Reward hacking creates unintended side effects. AI needs to follow social norms. Instrumental convergence: All intelligent agents will pursue common subgoals like the need for more power. t Goal Alignmen s Objectiv ? es and Rules Intelligent User Agent Side nt effe Alignme ct Actio n tra ra n ? prog i m Goal s select Dat Developers / Owners a Credit: Terminator 3: Rise of the Machines. Warner B Outlook AI is a technology that is on the verge of significant leaps… New technologies always had a profound impacted on the way we live and work (e.g., electricity, the internet, mobile communication). We can expect unprecedented gains in productivity from better narrow AI. New technologies always also present dangers and need to be regulated. This course will introduce simple techniques to create intelligent agents.