Introduction
SciFi x AI
The perception of AI changed over the decades. This is well reflected in Sci-Fi Movies, e.g.
- R2D2 in Star Wars
- The Terminator
- Agent Smith in the Matrix
- Sonny in I,Robot
- Wall·E
- etc.
Quotes
Will the fictions become reality? We will see. In any case, the intellectual frontiers of AI are wide open. The subfields span from artificial general intelligence (learning, reasoning, perception, etc.) to specific fields (e.g., translating, playing go) (Russel and Norvig 2022).
I believe it’s going to change the world more than anything in the history of mankind — even more than electricity. Kai-Fu Lee
The pace of progress in artificial intelligence is incredibly fast. Unless you have direct exposure to groups like Deepmind, you have no idea how fast—it is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five-year timeframe. 10 years at most. Elon Musk
Forget artificial intelligence—in the brave new world of big data, it’s artificial idiocy we should be looking out for. Tom Chatfield
Terms & definitions
What is AI?
AI is the science of making machines (i.e., computer systems) to think and to act humanly and/or rationally
AI tries to integrate cognitive abilities, such as inference, planning, problem solving, learning etc. in computer systems (i.e., think).
The goal of AI research is to enable machines to autonomously process and solve abstract tasks and problems even under changed conditions — without a human being programming an explicit solution (i.e., act).
All AI systems that are technically feasible today enable problem solving in limited contexts (e.g. speech or image recognition) and thus belong to the so-called weak AI.
The term rational is used here 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
Think humanly
Cognitive science is the study of the human brain and its processes — it examines how the human brain may be functioning. Cognitive science requires analytical observation and experimentation.
We can learn about human thought in three ways (Russel and Norvig 2022):
- introspection (trying to catch our own thoughts as they go by)
- experiments (observing a person in action)
- brain imaging (observing the brain in action)
As a result of the cognitive modelling approach, some of the most powerful AI models are a result from observing human thinking experimentally (e.g., deep neural networks).
Cognitive science is about making computers solve complex problems similar to how humans solve problems. Cognitive computing tries to replicate how humans would solve problems, while AI is not bound to human cognitive processes.
Think rationally
The laws of thought refer to fundamental axiomatic rules upon which rational discorse itself is often considered to be based.
Socrates is a man and all men are mortal, thus, it can be concluded that Socrates is mortal Aristotle (384-322 BCE)
Computers have been able to solve any solvable problem, as long as
- there are statements about any objects in the world,
- statements about the relations among them, and
- there is sufficient computing power available
These laws of thought were supposed to govern the operation of the mind; the studies initiated the field called logic. However, human decisions are not always mathematically perfect or logical.
Logic as conventionally understood requires knowledge of the world that is certain. As this condition is seldom achieved, the theory of probability fills this gap. Probability allows for rigorous reasoning with uncertain information. (Russel and Norvig 2022)
Act humanly
The Turing Test1 (Turing 1950) tests if a computer has the ability to mimic peoples’ behavior.
To pass the test, it would need following capabilities:
- natural language processing (communicate)
- knowledge representation (store information)
- automated reasoning (answer questions, draw new conclusions)
- machine learning (adapt to new circumstances)
Want to do a Turing Test? Play “Bot or Not”
The Turing Test (Turing 1950) was designed as a thought experiment that would sidestep the philosophical vagueness of the question “Can a machine think?” A computer passes the test if a human interrogator, after posing some written questions, cannot tell whether the written responses come from a person or from a computer (Russel and Norvig 2022).
The Total Turing Test additionally requires interaction with objects and people in the real world. This also requires computer vision and robotic capabilities.
Act rationally
An agent is something that acts, an rational agent is one that acts so as to achieve the best coutcome (i.e., does the right thing), or, when there is uncertainty, the best expected outcome (i.e., does the appropriate thing) based on the objective that is provided to the agent (Russel and Norvig 2022).
The approach goes beyond the “laws of thought” approach as it involves actions based on
- inference (deducing that a given action is the best and then to act on this conclusion) and
- other mechanisms such as reflex (when speed is more successful than careful deliberation that takes some time)
Russel and Norvig (2022) call the approach where the primary definition of success is getting better and better at achieving rigid human-specified goals the standard model of AI research. This standard model prevails not only in AI, “but also in control theory, where a controller minimizes a cost function; in operations research, where a policy maximizes a sum of rewards; in statistics, where a decision rule minimizes a loss function; and in economics, where a decision maker maximizes utility or some measure of social welfare” (p. 22).
They also criticize the standard model because it is increasingly difficult to specify the goal completely and correctly (e.g., autonomous driving involves multiple goals such as reaching the goal safely, where a strict definition of safety requires staying in the garage because driving on the road has a risk of injury due to myriad factors; how should the trade-off between reaching the goal and taking a risk of injury be made? There are so many questions that are difficult to answer a priori). Mis-specified goals most likely do not reflect what human designers intend, e.g., by not taking into account human values that are not included in the goals.
Benificial machines
To create machines that are provably beneficial to humans, two refinements to the standard model of AI are needed:
- The ability of any agent to choose rational actions is constrained by the computational untractability of doing so (Russel and Norvig 2022).
- An intelligent agent should not pursue a definite object, it should pursue objectives that benefit humans, while being uncertain as to what they are (Russel and Norvig 2022).
When a machine knows that it does not know the entire objective, it has an incentive to act cautiously, to ask permission, to learn more about our preferences through observation, and to submit to human control.
Definition of the EU
‘AI system’ means a machine-based systems designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers, from the input it received, how to generate output such as content, predictions, recommendations, or decisions, that can influence physical or virtual environment (European Commission 2024).
The history of AI
The Thinking Machine
A series of interviews to some of the AI pioneers.
The full documentary is available here
A brief AI-timeline
1943—1956 The inception of AI
- 1943: McCulloch & Pitts: Boolean circuit model of brain (artificial neurons with on and off states; all logical connectives can be implemented with some network of these)
- 1950: Turing’s “Computing Machinery and Intelligence” (Turing (1950) already introduced the Turing test, machine learning, genetic algorithms, and reinforcement learning)
- 1950s: Early AI programs (e.g., Arthur Samuel’s influential checkers program that learned to play at a strong amateur level)
1966—73 A dose of reality (AI winter)
- The early AI programs failed on more difficult problems
- Focus on “informed introspection” as to how humans perform a task
- Lack of appreciation of the intractability of many of the problems
- Signification reduction of government funding of AI research
1970—90 Expert systems (knowledge-based approaches)
- 1969—79: Early development of knowledge-based systems (rule-based heuristic algorithms)
- 1980—88: Expert systems industry booms (nmany U.S. corporates had their own AI groups)
- Soon after that came the “AI winter” (difficulties to build expert systems for complex domains due to uncertainty and a lack of learning)
1990—present AI spring (statistical approaches)
- Focus on probabilistic reasoning (rather than Boolean logic) and machine learning
- Reunification of subfields such as computer vision, robotics, speech recognition, and natural language processing
2012—present New excitement
- Advances in computing power, WWW, and very large data sets
(e.g., IBM Watson’s victory in Jeopardy!) - AI is accessible to many as it enters productivity tools (e.g., ChatGPT, Microsoft Co-Pilot)
New excitement
Improvement in performance obtained from increasing the size of the data set by two or three orders of magnitude outweighs any improvement that can be obtained from tweaking the algorithm Banko and Brill (2001)
- Deep learning systems offer significant performance gains
(e.g., AlphaGo’s victories) - Significant focus on AI in academia and industry
- Breakthrough of generative AI (e.g., ChatGPT)
- AI systems find increasing application in the real world (e.g., robotic vehicles, machine translation, speech recognition, recommendations, autonomous planning, game playing, image understanding, medicine)
Exercises
The exercises are (inspired) by Russel and Norvig (2022)
Concepts
Define in your own words:
- intelligence
- artificial intelligence
- agent
- rationality
- logical reasoning
Instances of AI
If and to what extent are the following computer systems instances of artificial intelligence?
- Supermarket bar code scanners
- Web search engines
- Voice-activated telephone menus
- Internet routing algorithms that respond dynamically to the state of the network
AI Contests
Various subfields of AI have held contests by defining a standard task and inviting researchers to do their best. Examples include the DARPA Grand Challenge for robotic cars, the International Planning Competition, the Robocup robotic soccer league, the TREC information retrieval event, and contests in machine translation and speech recognition.
Investigate one of these contests and describe the progress made over the years.
- To what degree have the contests advanced the state of the art in AI?
- To what degree do they hurt the field by drawing energy away from new ideas?
Statements
Read the statements (one after the other) and discuss if the second sentence of each statement is true and if it does imply the first.
Surely computers cannot be intelligent
—they can do only what their programmers tell them.
Surely animals cannot be intelligent
—they can do only what their genes tell them.
Surely animals, humans, and computers cannot be intelligent
—they can do only what their constituent atoms are told to do by the laws of physics.