CMP 411: Artificial Intelligence

1. Definition of Artificial Intelligence (AI)

AI refers to the simulation of human intelligence in machines that are programmed to think, learn, and
make decisions like humans. These systems are designed to perform tasks that typically require human
intelligence, such as visual perception, speech recognition, and decision-making,

In their seminal textbook "Artificial Intelligence: A Modern Approach," Stuart Russell and Peter Norvig
describe four distinct definitions or approaches to Artificial Intelligence (AI). These definitions are based
on two key dimensions:

 Thinking vs. Acting – Whether the focus is on thought processes and reasoning or on behavior
and actions.
 Human-like vs. Rational – Whether the goal is to mimic human intelligence or to achieve optimal,
rational outcomes.

Combining these dimensions, Russell and Norvig outline four possible definitions of AI as follows:

Thinking Like Human (Cognitive Modeling): This approach focuses on creating machines that think like
humans. It involves modeling human cognition and replicating human thought processes. Here, the goal
is to understand how humans think and then design AI systems that mimic those processes. This often
involves interdisciplinary work with cognitive science, psychology, and neuroscience. Example include
developing AI systems that solve problems in the same way humans do, such as using heuristics or mental
models.

Acting Like Human (Turing Test Approach): This approach focuses on creating machines that act like
humans. The goal is to build systems that can perform tasks in a way indistinguishable from human
behavior. The Turing Test, proposed by Alan Turing, is a benchmark for this approach. If a machine can
interact with a human in such a way that the human cannot tell whether they are interacting with a
machine or another human, the machine is considered intelligent. Example systems within this definition
include Chatbots or virtual assistants that can hold natural conversations with humans.

Thinking Rationally (Laws of Thought Approach): This approach focuses on creating machines that think
rationally, using logical reasoning to solve problems. It is based on formal logic and mathematical
principles. The goal is to design systems that use logical rules to arrive at correct or optimal conclusions.
This approach assumes that intelligence is rooted in the ability to reason logically. Example of this
approach are expert systems that use rule-based reasoning to diagnose medical conditions or solve
mathematical problems.

Acting Rationally (Rational Agent Approach): This approach focuses on creating machines that act
rationally, meaning they take actions that maximize their chances of achieving their goals, given the
available information. The emphasis is on practical decision-making and behavior rather than mimicking
human thought. A rational agent is one that acts to achieve the best possible outcome in its environment.
Examples within this definition are the self-driving cars that make decisions to navigate safely and
efficiently, or recommendation systems that optimize user satisfaction.

1
Russell and Norvig argue that the rational agent approach is the most comprehensive and practical
definition of AI, as it encompasses both reasoning and action while focusing on achieving optimal
outcomes in real-world environments. This approach has become the foundation for much of modern AI
research and development.

2. Strong AI Vs Weak AI

Russell and Norvig further classifies systems associated with thought (thinking like humans or thinking
rationally) as Strong AI. Systems that are associated with intelligent actions only (Acting like humans or
acting rationally) are classified as weak AI.

3. Mundane vs. Expert Tasks

Mundane Tasks

These are everyday tasks that people perform, that do not require specialized knowledge or deep
expertise. They can usually be completed with minimal training.

Examples of Mundane Tasks:

 Vision – Ability to make sense of what we see.

 Natural Language – ability to hold conversations in natural languages.
 Robotics – Ability to move around freely without hitting objects
 Objects/face recognition – Ability to recognize objects and people we know.
 Voice – Ability to recognize voices of people we know

Expert Tasks

These are complex activities that require specialized knowledge, critical thinking, creativity, or extensive
experience. Completing them effectively often requires training, education, or hands-on experience.

Examples of Expert Tasks:

 Software Development – Writing and debugging code for applications.

 Legal Consulting – Providing legal advice based on laws and precedents.
 Medical Diagnosis – Analyzing patient symptoms and test results to identify illnesses.
 Data Science & Machine Learning – Analyzing large datasets and building predictive models.
 Aircraft Maintenance & Repair – Troubleshooting and fixing complex aviation systems.

AI aims to automate both mundane and expert tasks. It turns out that mundane tasks are by far more
difficult to automate even as they are easy for humans. Expert tasks, on the other hand, are easier to
automate even as they tend to be more difficult for humans.

4. Goals of Artificial Intelligence

The goals of Artificial Intelligence (AI) can be broadly categorized into scientific goals and engineering
goals.

Scientific Goals of AI

2
The scientific goals of AI are rooted in understanding and advancing knowledge about intelligence,
cognition, and problem-solving. These goals are primarily driven by curiosity and the desire to explain
natural phenomena. The following are some scientific goals of AI

Understanding Intelligence: The primary scientific goal of AI is to understand the nature of intelligence,
both human and artificial. This involves studying how humans and animals think, learn, reason, and make
decisions. Under this goals, researchers aim to uncover the principles underlying intelligence, such as
perception, reasoning, memory, and decision-making.

Modeling Cognitive Processes: In this goal, AI scientists seek to create computational models that
replicate or simulate human cognitive processes. This includes areas like natural language understanding,
vision, and problem-solving. For example, neural networks are inspired by the structure and function of
the human brain, and their development is partly motivated by the desire to understand biological
intelligence.

Exploring Theoretical Limits: Scientific AI research investigates the theoretical boundaries of what
machines can and cannot do. This includes questions about computability, complexity, and the limits of
learning algorithms. For instance, research in computational theory explores whether certain problems
are inherently unsolvable by machines.

Interdisciplinary Insights: AI science often intersects with other disciplines, such as neuroscience,
psychology, linguistics, and philosophy, to gain a deeper understanding of intelligence. For example,
cognitive science uses AI models to test hypotheses about how the human mind works.

Advancing Fundamental Knowledge: Scientific AI focuses on discovering new algorithms, theories, and
frameworks that expand our understanding of intelligence and computation. This includes developing
new learning paradigms, such as reinforcement learning or unsupervised learning.

Engineering Goals of AI

The engineering goals of AI are focused on building practical, efficient, and reliable systems that solve
real-world problems. These goals are driven by the need to create useful applications and technologies.
They include the following:

Building Functional Systems: The primary engineering goal of AI is to design and implement systems that
perform specific tasks effectively. This includes applications like speech recognition, image classification,
autonomous vehicles, and recommendation systems. Engineers prioritize functionality, scalability, and
usability over theoretical elegance.

Optimizing Performance: AI engineers aim to optimize systems for accuracy, speed, and resource
efficiency. This involves fine-tuning algorithms, improving hardware-software integration, and reducing
computational costs. For example, optimizing deep learning models for faster inference on mobile devices
is a key engineering challenge.

Solving Real-World Problems: Engineering AI focuses on addressing practical challenges in industries such
as healthcare, finance, transportation, and manufacturing. Examples include developing AI-powered
diagnostic tools, fraud detection systems, and supply chain optimization algorithms.

3
Ensuring Robustness and Reliability: AI engineers strive to create systems that are robust, reliable, and
safe. This includes addressing issues like bias, adversarial attacks, and failure modes. For instance,
ensuring that self-driving cars can handle unexpected scenarios is a critical engineering goal.

Scaling and Deployment: Engineering AI involves scaling solutions to handle large datasets, high user
demand, and diverse environments. This requires expertise in software engineering, cloud computing,
and distributed systems. For example, deploying AI models in production environments with continuous
updates and monitoring is a key engineering task.

User-Centric Design: AI engineers focus on creating systems that are intuitive and accessible to users. This
involves designing user interfaces, ensuring explainability, and addressing ethical concerns. For instance,
developing AI systems that provide transparent explanations for their decisions is an important
engineering goal.

Table 1.1: Key Differences between Scientific and Engineering Goals

Aspect Scientific Goals Engineering Goals

Primary Focus Understanding intelligence and Building practical, functional systems

cognition

Motivation Curiosity-driven, knowledge Problem-solving, application

advancement development

Outcome Theories, models, and insights Products, tools, and solutions

Methodology Hypothesis testing, experimentation Design, optimization, and deployment

Metrics of Success Novelty, explanatory power Accuracy, efficiency, reliability

Interdisciplinary Neuroscience, psychology, philosophy Software engineering, robotics,

Links industry

Interplay Between Scientific and Engineering Goals

While the scientific and engineering goals of AI are distinct, they are deeply interconnected. On one hand,
science informs engineering i.e. scientific discoveries, such as new learning algorithms or cognitive
models, often lead to breakthroughs in engineering applications. On the other hand, it is the engineering
challenges that drive science i.e. practical problems encountered in engineering can inspire new scientific
research. For example, the need for explainable AI has spurred research into interpretable machine
learning models. Many AI projects involve addressing both scientific and engineering goals to advance
knowledge and create impactful technologies.

5. Philosophical Aspects of AI

The idea of machines being autonomously intelligent is a profound one which did not go unchallenged
right from its inception. It raises fundamental questions about the nature of intelligence, consciousness,
ethics, and the relationship between humans and machines. These questions have been debated for
decades and continue to evolve as AI technologies advance.

4
The Turing Test

Alan Turing, a pioneering figure in computer science and artificial intelligence, introduced the concept of
the Turing Test in his 1950 paper "Computing Machinery and Intelligence." The Turing Test is designed to
evaluate a machine's ability to exhibit intelligent behavior indistinguishable from that of a human. The
Turing Test is structured as follows:

1. Setup:

 A human judge interacts with two entities: a machine (AI) and a human.
 The interaction is text-based (e.g., via a computer terminal), and the judge cannot see or
hear the entities.

2. Objective:

The judge asks questions or engages in conversation with both the machine and the
human.
 The machine's goal is to respond in a way that convinces the judge it is the human, while
the human responds naturally.
3. Outcome:

o If the judge cannot reliably distinguish the machine from the human based on their
responses, the machine is said to have passed the Turing Test.

Figure. 1.1: Illustration of the Turing Test

The "standard interpretation" of the Turing test, in which human C, the interrogator, is given the task of
trying to determine which entity – A or B – is a computer and which is a human. The interrogator is
limited to using the responses to written questions to make the determination.

The Chinese Room Experiment

5
The Chinese Room argument is a thought experiment put forward by John Searle which challenges the
idea of strong AI by demonstrating that even if a computer can convincingly simulate understanding
language (i.e pass the Turing Test), it doesn't truly understand it.

Here's a breakdown:

 The Setup:

Imagine a person who doesn't know Chinese is locked in a room with a set of Chinese characters and a
rulebook (written in English) that specifies how to manipulate those characters.

 The Interaction:

Someone outside the room sends Chinese questions or messages through a slot, and the person inside
uses the rulebook to generate appropriate Chinese responses, which are then sent back out.

 The Point:

Even if the person in the room can produce fluent and coherent Chinese responses, they don't
understand the meaning of the characters they are manipulating; they are merely following rules.

 The Argument:

Searle argues that this demonstrates that a computer program, which is essentially a set of rules, can
simulate understanding language, but it doesn't actually understand it in the same way a human does.