Lecture Notes on Artificial Intelligence - Introduction
Introduction
Course Opening and Welcome
The lecture commences with a formal welcome to the Artificial Intelligence course, designed for the Master’s programs in Computer Science, Artificial Intelligence and Cyber Security, and CMTI. Professors Stefano Mizzaro and Giuseppe Serra are introduced as the co-instructors, who will jointly deliver the course content, dividing the lectures equally.
Instructors and Course Overview
Instructors
Stefano Mizzaro: Professor Mizzaro will lead the first half of the course, amounting to approximately 24 hours of instruction. His contact email is provided for student inquiries. His primary research interests encompass the evaluation of Information Retrieval systems, including search engines like Google, crowdsourcing methodologies, the study of disinformation and fake news, and the development of context-aware mobile applications. With 32 years of experience in computer science, Professor Mizzaro characterizes his approach as focusing on the interdisciplinary aspects of informatics, exploring the connections and applications of computer science in broader fields, describing himself humorously as a "failed computer scientist" due to his inclination towards the periphery of core computer science rather than its central theoretical aspects.
Giuseppe Serra: Professor Serra will conduct the second half of the course, also contributing approximately 24 hours of lectures. His teaching will concentrate on the practical applications of Artificial Intelligence, particularly within machine learning and deep learning paradigms. A significant focus will be placed on reinforcement learning, a domain of considerable contemporary relevance, exemplified by its application in advanced AI systems such as chat GPT. Professor Serra aims to provide students with a comprehensive understanding of these cutting-edge techniques and their real-world impact.
Course Objectives
The principal objectives of this Artificial Intelligence course are:
Defining AI: To rigorously introduce and explore the concept of Artificial Intelligence, addressing the inherent complexities and varied interpretations of its definition.
Scope of AI: To delineate the extensive scope of the field of AI and to specify the topics that will be covered within this particular course, acknowledging the breadth of AI as a dynamic and multifaceted discipline.
Course Operations: To furnish students with essential operational details concerning the course structure, learning materials, and methods of assessment, ensuring clarity and preparedness.
Motivation for Studying AI: To articulate the compelling reasons for studying Artificial Intelligence, emphasizing its growing importance and diverse applications across numerous sectors.
Historical and Current Context: To provide a concise overview of the historical evolution of AI and to contextualize its current state, setting the stage for deeper explorations in subsequent lectures.
Course Logistics and Practical Information
Schedule, Timings, and Attendance
Lecture Schedule: Classes are scheduled to take place on Mondays and Tuesdays, specifically during the lunchtime period to accommodate various student schedules.
Instructor Allocation: The course is divided into two main segments. Professor Stefano Mizzaro is slated to conduct the initial ten lectures, focusing on the foundational aspects of AI. Professor Giuseppe Serra will then take over for the subsequent lectures, delving into more specialized topics. The precise distribution of the final lectures will be determined as the course progresses, considering both instructors’ commitments and the course’s evolving needs.
Adjusted Lecture Timing: To better accommodate students transitioning from prior classes and to mitigate tardiness, the lecture start time has been adjusted to 12:40 PM. This slight delay from the originally planned 12:30 PM start aims to ensure that students can arrive punctually and settle in before the lecture begins. Consequently, the lectures are expected to conclude around 2:25 PM. This revised schedule will be evaluated after the next lecture to ensure it effectively meets the needs of the students and the course. The rationale behind this adjustment is to maximize attendance from the beginning of each lecture, fostering a more cohesive and less disruptive learning environment.
Lecture Breaks: While traditionally a short break is incorporated mid-lecture, it is anticipated that for the initial weeks, lectures will proceed without scheduled breaks. This is partly to compensate for the slightly later start time and to maintain momentum in covering the introductory material. However, the instructors remain flexible and may reintroduce breaks as deemed necessary or beneficial for student engagement and comprehension as the course progresses. Any changes to this policy will be communicated in advance.
Learning Resources and Materials
Lecture Slides and Recordings (Teams Platform)
Microsoft Teams as Primary Platform: All essential course materials, including lecture slides and recordings, will be hosted on the Microsoft Teams platform. Students are expected to join the designated course Team to access these resources and receive course-related announcements.
Pre-Lecture Slide Availability: Lecture slides are typically uploaded to the Teams group prior to each lecture. This is intended to facilitate proactive learning, allowing students to review the material in advance and come prepared with questions. Furthermore, students who prefer to take notes directly on the slides can download them beforehand and annotate during the lecture.
Comprehensive Lecture Recordings: Efforts will be made to record each lecture session and make these recordings available on Teams. Recordings will be attempted using both Microsoft Teams’ built-in recording feature and OBS (Open Broadcaster Software) to ensure redundancy and potentially higher quality. However, students should be aware that technical issues may occasionally affect the availability or quality of recordings. These recordings serve as a valuable resource for review, catching up on missed sessions, or clarifying complex topics discussed in class.
Slide Updates and Official Versions: While slides are made available before lectures for convenience, students should note that these are preliminary versions. Instructors reserve the right to modify and refine the slides during or after the lecture to correct errors, enhance clarity, or incorporate new insights. The most current and official versions of the slides are those available on Teams post-lecture. Students are not individually notified of slide updates; it is the student’s responsibility to ensure they are using the latest version from Teams for study purposes.
Textbook: "Artificial Intelligence: A Modern Approach"
Primary Textbook: The cornerstone textbook for this course is "Artificial Intelligence: A Modern Approach," authored by Stuart Russell and Peter Norvig. Widely recognized as a definitive and encyclopedic resource in the field, it provides a comprehensive and in-depth exploration of AI principles, techniques, and applications. It is used in the majority of AI courses globally and is considered a standard reference.
Availability in English and Italian: The textbook is available in both English (Global Edition) and Italian. The Italian version is a well-regarded translation that has been adapted into two volumes for practical reasons, whereas the English version is typically a single, more substantial volume. Students may choose either version based on preference, as the core content is consistent across both. The Global Edition in English is specified as the reference, noting minor variations in chapter order across different editions.
Textbook as a Guide, Not Sole Source: Lectures will be structured to align with and reference specific chapters or sections of the textbook. However, the textbook serves as a guide and a deeper resource rather than the sole source of information. Some topics will be covered in detail as presented in the book, while others may be selectively addressed or supplemented with additional materials. In certain instances, lectures may diverge from the textbook to explore contemporary research, alternative perspectives, or topics not extensively covered in the book. Instructors will clearly indicate the relevant textbook sections for each lecture to facilitate student reading and preparation.
Comprehensive yet Potentially Advanced Nature: "Artificial Intelligence: A Modern Approach" is exceptionally comprehensive, aiming to cover a vast spectrum of AI. Its depth and technical detail make it almost encyclopedic. While this thoroughness is a strength, it also means that certain sections are quite advanced, potentially more suited to a doctoral-level course. For a Master’s level course, some parts might delve into unnecessary technical depth. The instructors are aware of this and will present the material in a manner that is accessible and appropriately leveled for Master’s students. The aim is to convey the essential concepts effectively, enabling students to engage with the textbook for deeper understanding while focusing on the most pertinent aspects for the course objectives. Students are encouraged to use the textbook to reinforce lecture material and explore topics further, with the understanding that some of the more technically intricate sections can be approached selectively.
Assessment and Examination
Proposed Exam Dates and Sessions
Tentative Exam Schedule: The proposed examination dates for the initial examination session are set for July 5th, July 22nd, and September 26th. It is crucial to note that these dates are currently provisional and are pending formal approval from the academic commission. Students should consider these dates as indicative and await official confirmation, which will be announced through course communication channels once received.
Full-Day Exam Sessions: Examinations are planned to be conducted across full-day sessions to accommodate both written and oral components. The intention is to administer the written exam in the morning and, where feasible, conduct the oral exams in the afternoon of the same day. This format is designed to streamline the examination process and provide timely feedback. However, the practicality of conducting both exam components on the same day is contingent upon the number of students participating in each exam session.
Contingency for Large Classes: In scenarios where a large number of students are registered for an exam session, it may not be logistically possible to complete both the written exam correction and all oral exams within a single day. In such cases, the written exam will still be held in the morning, but the oral exams will be scheduled for one or more subsequent days immediately following the written exam. This adjustment ensures that each student receives adequate time and attention during their oral examination, and that the grading process remains thorough and fair, even with a large cohort. The exact scheduling of oral exams will be communicated to students after the written exam, based on the number of students who have taken the written component and the logistical constraints.
Exam Structure: Written and Oral Components
Dual-Component Assessment: The final course assessment comprises two equally weighted components: a written examination and an oral examination. Each component contributes approximately 50% to the overall final grade, emphasizing the importance of both written comprehension and verbal articulation of the course material.
Written Exam as Qualifying Filter: The written exam serves as a critical filter for the oral component. Successful completion of the written exam is a prerequisite for eligibility to proceed to the oral examination. Failing the written exam will preclude a student from taking the oral exam and, consequently, from passing the course. This structure ensures that students demonstrate a foundational understanding of the material in writing before being assessed on their deeper comprehension and ability to discuss the topics orally.
Written Exam Format and Content: The written examination is designed to be completed within approximately two hours. The content of the exam will be evenly distributed, with roughly half of the questions pertaining to the material covered by Professor Mizzaro and the other half focusing on topics taught by Professor Serra. Questions will primarily be open-ended, requiring students to demonstrate problem-solving skills, critical thinking, and conceptual understanding. The exam will not employ multiple-choice questions or similar formats that test rote memorization. Instead, it will emphasize the application of knowledge and the ability to articulate reasoned responses to complex prompts, such as defining AI concepts, explaining techniques, or solving problems using AI methodologies discussed in the course.
Oral Exam for Grade Adjustment and In-Depth Evaluation: The oral examination provides an opportunity for a more nuanced and in-depth evaluation of a student’s understanding. It is not merely a formality but a significant component that can substantially modify the grade initially indicated by the written exam. The oral exam can lead to both positive and negative adjustments to the written exam score. A strong performance in the oral exam, demonstrating deep comprehension, insightful analysis, and the ability to engage critically with the material, can significantly повысить the final grade, even beyond what the written exam score might suggest. Conversely, a poor performance in the oral exam, revealing misunderstandings or superficial knowledge, can lead to a downward revision of the grade. This dynamic interaction between written and oral assessment ensures a holistic and accurate evaluation of each student’s mastery of the course content. Examples of typical exam questions will be provided to students in due course to aid in their preparation, although the specific questions in the actual exam will be novel and designed to assess genuine understanding rather than rote learning.
Student Background and Introductory Questionnaire
Purpose of the Questionnaire: To effectively tailor the course content and delivery, an introductory questionnaire has been prepared to gather essential information about the students’ academic backgrounds. This questionnaire aims to create a clearer profile of the class composition, including prior degrees, program affiliations, and relevant prior coursework.
Information Collected: The questionnaire seeks to collect specific details such as:
Prior Degree and Institution: To understand the academic foundation students are building upon.
Current Program of Study: To ascertain whether students are enrolled in Computer Science, Artificial Intelligence and Cyber Security, CMTI, or other programs, as this indicates their primary academic focus and expected learning outcomes.
Specialization or Track within Program: For programs with specializations (e.g., within Computer Science or IoT Big Data and Machine Learning), to identify specific areas of concentration that may influence students’ perspectives and interests in AI.
Relevant Coursework: To gauge the level of prior exposure to subjects directly relevant to AI, such as logic, probability, statistics, algorithms, machine learning, and related areas. This helps in understanding the baseline knowledge of the class.
While the questionnaire asks for name and surname for organizational purposes, students are assured that if they prefer to use a pseudonym, it will not affect their participation or assessment. The primary focus is on the background information that will aid in course customization.
Utilizing Background Information for Course Customization: The data gathered from the questionnaire is crucial for informing the instructors about the diverse academic profiles within the class. For instance, recognizing that some students may not have a formal background in logic (common in Computer Science degrees but less so in others like IoT Big Data and Machine Learning) allows the instructors to adjust the level of assumed prior knowledge when discussing logic-based AI approaches. Similarly, understanding the distribution of students from different programs helps in highlighting connections to various application domains relevant to their respective fields. By gaining a clearer picture of the students’ collective background, the instructors can make informed decisions about course pacing, depth of coverage for certain topics, and the inclusion of supplementary materials or examples that cater to the specific needs and knowledge levels of the students. This personalized approach aims to maximize the learning experience and ensure that the course is both challenging and accessible to all students, regardless of their precise academic trajectory leading up to this course.
Defining Artificial Intelligence: Initial Concepts
Brainstorming and Eliciting Initial Definitions of AI
The lecture initiates the exploration of Artificial Intelligence by engaging students in a brainstorming session to elicit their preliminary understandings and definitions of AI. This interactive approach serves to uncover pre-existing notions and set the stage for a more formal and nuanced discussion. The initial suggestions from the students included a range of insightful perspectives:
Simulating Human Behavior: This suggestion highlights the common perception of AI as systems that mimic human actions or reactions. It touches upon the Turing Test concept, where a machine’s intelligence is judged by its ability to convincingly imitate a human in conversation. This perspective is rooted in the idea that intelligence is observable through behavior and that replicating human behavior is a key goal of AI.
Taking Non-Deterministic Decisions and Exhibiting Non-Deterministic Behavior: These related points emphasize the capacity of AI systems to make choices or act in ways that are not strictly predetermined. Unlike traditional computer programs that follow a fixed set of rules, AI, particularly modern systems incorporating machine learning, can exhibit variability and unpredictability in their outputs. This non-deterministic aspect is crucial for tasks requiring creativity, adaptation, and handling uncertainty, mirroring human decision-making processes which are often not purely algorithmic. For instance, in generative models like chat GPT, the same prompt can yield different, yet relevant and coherent, responses each time, showcasing non-deterministic behavior.
Systems that Solve Problems without Explicit Programming: This definition gets to the heart of machine learning and modern AI. It points out that AI systems are not merely executing pre-written algorithms for specific problems but are designed to learn from data and generalize to solve new, unseen problems. This contrasts with traditional software engineering where solutions are explicitly coded. In AI, especially with machine learning, the system learns the ‘how’ of problem-solving from data, rather than being explicitly told step-by-step instructions.
Solving Non-Algorithmic Problems: This suggestion delves deeper into the nature of problems AI can address. It suggests that AI is capable of tackling problems that are not easily solvable by traditional algorithms, either because the algorithms are unknown, too complex to formulate, or because the problem itself is ill-defined or requires subjective judgment. Examples of non-algorithmic problems include image recognition, natural language understanding, and decision-making in complex, dynamic environments. These are tasks where human intuition and learning play a significant role, and AI aims to replicate or even surpass these human capabilities.
Collectively, these initial student ideas effectively capture several fundamental aspects of Artificial Intelligence. They touch upon the themes of mimicking human intelligence, dealing with uncertainty and variability, learning from data, and tackling complex problems that defy traditional algorithmic solutions. These initial definitions provide a solid foundation for the subsequent, more structured exploration of what constitutes AI.
Exploring External Perspectives and Media Representations
The lecture then shifts to examining how Artificial Intelligence is commonly portrayed in external sources, particularly in media and popular culture. This exploration is crucial because public perception significantly shapes understanding and expectations of AI, often diverging from the technical realities and scientific discourse. The discussion highlights several key trends in media representations:
Films: Apocalyptic Scenarios and Futuristic Visions: Cinema frequently depicts AI in dramatic and often dystopian contexts. A prevalent theme is the apocalyptic scenario where AI, often embodied as robots, becomes a threat to humanity. Films like "The Terminator" or "The Matrix" exemplify this, portraying AI as antagonistic forces seeking to dominate or destroy human civilization. Conversely, other films present futuristic, sometimes utopian, sometimes strange visions of AI integrated into society, often exploring the ethical and existential implications of advanced AI. While robots are a common visual representation of AI in films, it’s important to note that AI is not inherently tied to physical robots. However, cinematic portrayals often anthropomorphize AI, focusing on humanoid robots as the embodiment of artificial intelligence. An exception mentioned is HAL 9000 from "2001: A Space Odyssey," which represents AI as a computational system, a computer, rather than a physical robot, highlighting that AI can exist in non-corporeal forms. These cinematic representations, while often sensationalized, contribute to public consciousness and shape initial perceptions of AI’s potential and perils.
News: Sensationalism, Investment, and Societal Impact: News media often focuses on the more sensational and alarming aspects of AI to capture audience attention. Headlines frequently highlight potential risks, job displacement fears, and ethical dilemmas associated with AI advancements. Apocalyptic scenarios, mirroring film tropes, sometimes find their way into news narratives, albeit often framed as potential future risks rather than current realities. However, news coverage also acknowledges and reports on the massive financial investments pouring into AI research and development, recognizing AI as a major technological and economic force. Significant achievements in AI, such as breakthroughs in natural language processing (like chat GPT), image recognition, and AI’s application in various industries, are also reported, though often alongside discussions of their societal impact. The lecture points to the Cambridge Analytica scandal as an example of the societal impact of AI when combined with social media, illustrating the potential for misuse and ethical concerns around data privacy and algorithmic bias. Furthermore, the emergence of regulatory responses like GDPR (General Data Protection Regulation) and the AI Act are mentioned, indicating a growing societal and governmental awareness of the need to govern AI development and deployment to mitigate risks and ensure responsible innovation.
Hype and Scientific Claims: Exaggeration and Unfulfilled Promises: A critical aspect discussed is the presence of significant hype surrounding AI, often fueled not just by media but also by AI scientists and researchers themselves. Historically, there have been instances of over-optimistic predictions and exaggerated claims about the near-term capabilities of AI. The lecture references examples from the early days of AI research where bold pronouncements about achieving human-level AI within a few years were made, predictions that have not yet materialized decades later. This historical context is important to understand the cyclical nature of AI enthusiasm and periods of "AI winters" when initial hype gives way to disillusionment due to unmet expectations. It’s crucial to approach both media portrayals and even some scientific claims with a degree of skepticism and critical evaluation, recognizing the tendency towards exaggeration and the difference between current AI capabilities and long-term aspirations.
Despite the varied and often sensationalized perceptions of AI in media and popular culture, the lecture underscores that AI, even in its current state, is not merely hype. It is acknowledged that AI has achieved remarkable successes and demonstrated capabilities that were once considered science fiction. AI systems routinely perform at a superhuman level in specific, well-defined tasks. Examples provided include: medical diagnosis, where AI algorithms can now outperform human experts in detecting diseases from medical images; analysis of vast datasets from social media, enabling insights and predictions beyond human capacity; and game playing, where AI has conquered complex games like chess, Go, and even video games like Pac-Man, consistently surpassing human champions. These concrete examples serve to ground the discussion in the reality of AI’s current achievements, balancing the often exaggerated and sometimes misleading portrayals found in broader media.
Developing a Nuanced Understanding of AI Definition Challenges
To move beyond superficial understandings and towards a more robust definition of Artificial Intelligence, the lecture delves into several key nuances and challenges that complicate the very act of defining AI. These challenges arise from the multifaceted nature of intelligence itself and the diverse goals and approaches within the field of AI research.
Beyond Human Simulation: Superhuman Capabilities
The ambition of AI extends beyond simply mimicking or simulating human behavior. Modern AI increasingly aims to achieve capabilities that surpass human limitations in specific domains.
The lecture uses the example of AI excelling in games like chess and Go to illustrate this point. In these complex strategy games, modern AI programs, such as AlphaGo, have not only defeated the best human players but have also employed strategies and moves that were considered unconventional or even counter-intuitive from a human perspective. The famous "Move 37" in a game played by AlphaGo against Lee Sedol is often cited as an example of a move that no human Go player would likely have considered, yet it proved to be a pivotal, winning move. This demonstrates that AI is not merely imitating human gameplay but is exploring and exploiting the game space in ways that go beyond human intuition and strategic thinking.
AI systems achieve superhuman performance in games like Chess and Go, and in fields like medical diagnosis and financial analysis. For instance, AlphaGo’s "Move 37" in Go demonstrated a strategy beyond human intuition. In medical diagnosis, AI can exceed human radiologists in accuracy. In finance, AI algorithms analyze market data faster and more comprehensively than humans.
Therefore, defining AI solely in terms of human simulation is limiting. A more comprehensive definition must acknowledge AI’s capacity to not only replicate human intelligence but also to augment and surpass human abilities in specific, and increasingly numerous, domains. This raises questions about what constitutes "intelligence" in a broader sense, and whether AI should be evaluated based on human-centric benchmarks or on its ability to achieve optimal outcomes, regardless of whether those outcomes are achieved in a human-like manner.
Addressing Deterministic vs. Non-Deterministic Systems
Modern AI systems, particularly those based on machine learning, often exhibit non-deterministic behavior, unlike traditional deterministic computer programs.
The lecture highlights chat GPT as an example of a non-deterministic AI system. When prompted with the same question multiple times, chat GPT can generate different, yet contextually relevant and coherent, responses. This variability is not a flaw but a feature, stemming from the probabilistic nature of the underlying models and the complex processes involved in natural language generation. Unlike a calculator that will always give the same answer to "2+2," chat GPT’s responses are influenced by a multitude of factors, including the vast dataset it was trained on, the specific algorithms used, and even random elements introduced to enhance creativity and diversity in its outputs.
Chat GPT is an example of a non-deterministic AI system. It can generate different, yet relevant and coherent, responses to the same prompt due to its probabilistic nature and complex language generation processes.
This non-deterministic nature is not unique to large language models like chat GPT. Many AI systems, particularly those dealing with complex, real-world data or tasks, incorporate probabilistic models and learning algorithms that introduce variability. For instance, in reinforcement learning, an AI agent’s behavior evolves through trial and error, and even after learning, there can be stochastic elements in its decision-making process. In image recognition, while the goal is to consistently identify objects correctly, the internal processes of a neural network may involve non-deterministic operations, and the system’s response to slightly different or ambiguous inputs might vary.
The introduction of non-determinism in AI systems has profound implications for how we define and evaluateAI. It moves AI away from the realm of purely rule-based, predictable machines and closer to systems that exhibit adaptability, creativity, and a degree of unpredictability, traits often associated with biological intelligence. This necessitates definitions of AI that can encompass both deterministic and non-deterministic approaches, and that can account for systems whose behavior is not always fully predictable or explainable in terms of simple, fixed rules.
Problem Solving without Explicit Programming
A key feature of modern AI, especially machine learning, is its ability to solve problems without explicit programming for each instance, learning from data instead.
Traditional software development relies on programmers writing explicit, step-by-step instructions (algorithms) to solve specific problems. For every task a program is intended to perform, a programmer must anticipate the inputs, define the processing steps, and specify the outputs. This approach is effective for well-defined problems with clear algorithmic solutions, but it becomes increasingly challenging, and often infeasible, for complex, real-world problems where the rules are unknown, constantly changing, or too intricate to codify manually.
Machine learning, a core subfield of AI, offers an alternative paradigm. Instead of explicitly programming solutions, machine learning focuses on developing algorithms that enable systems to learn from data. In machine learning, an AI system is trained on a dataset of examples relevant to the problem it is intended to solve. For instance, to create an AI system that can recognize handwritten digits, a machine learning model is trained on a vast dataset of images of handwritten digits, along with the corresponding labels (0, 1, 2, ..., 9). The learning algorithm analyzes this data, identifies patterns and relationships, and automatically learns to map input images to the correct digit labels. Crucially, after training, the system can generalize its learning to recognize new, unseen handwritten digits, without needing to be explicitly programmed for each possible digit variation.
This ability to learn from data and generalize to new situations is a hallmark of intelligent behavior. It allows AI systems to tackle problems that are too complex or ill-defined for traditional programming approaches. Examples are abundant:
Image and Speech Recognition: It is practically impossible to write explicit rules to recognize all possible variations of objects in images or words in speech. Machine learning models, trained on vast datasets of images and audio, learn to perform these tasks with remarkable accuracy.
Natural Language Processing: Understanding and generating human language is incredibly complex due to its ambiguity, context-dependence, and vast vocabulary. Machine learning, particularly deep learning, has enabled significant breakthroughs in natural language processing tasks like machine translation, text summarization, and question answering.
Decision-Making in Complex Environments: For tasks like autonomous driving or robotic control in unstructured environments, the number of possible scenarios and contingencies is too large to pre-program. Reinforcement learning, a type of machine learning, allows AI agents to learn optimal strategies through trial and error, interacting with their environment and learning from feedback.
The capacity for problem-solving without explicit programming is not just a technical detail; it fundamentally changes the nature of computation and automation. It allows us to create systems that can adapt, learn, and solve problems in dynamic and unpredictable environments, opening up a vast range of applications that were previously unattainable with traditional computing methods. This capability is a cornerstone of modern AI and must be central to any comprehensive definition of the field.
Framing the Definition of AI: A Structured Approach
Moving Towards Formal and Expert Definitions
Having initiated our exploration of Artificial Intelligence through student-generated definitions and an examination of media portrayals, we now transition to a more structured and expert-driven approach. It is crucial to recognize from the outset that within the field of Artificial Intelligence, there exists no single, universally accepted definition. Instead, a variety of perspectives and approaches coexist, often leading to debates and differing research directions. This lack of a monolithic definition is not a weakness but rather reflects the richness and evolving nature of the field itself. To navigate this landscape of definitions, we turn to a well-established framework proposed by Russell and Norvig in their seminal textbook, "Artificial Intelligence: A Modern Approach." This framework provides a valuable lens through which to categorize and understand the diverse perspectives on what constitutes Artificial Intelligence.
Russell and Norvig’s Framework: Two Defining Dimensions
Russell and Norvig offer a structured approach to classifying definitions of AI by identifying two fundamental dimensions that differentiate various perspectives. These dimensions are not mutually exclusive but rather represent axes along which different definitions of AI can be positioned. Understanding these dimensions is key to appreciating the nuances and variations in how AI is conceptualized.
Focus on Thought vs. Action (Thinking vs. Acting)
Russell and Norvig’s first dimension for classifying AI definitions is the focus on either thought processes (thinking) or observable actions (acting).
Thinking Processes and Reasoning: This perspective centers on the internal cognitive mechanisms of AI systems. It is concerned with how AI systems process information, solve problems, reason logically, and make inferences. Definitions aligned with this dimension seek to model or replicate the internal workings of the human mind, or at least to create systems that exhibit rational thought, regardless of whether it mirrors human cognition. The focus is on the mental processes underlying intelligent behavior. For example, early AI research in symbolic AI, which aimed to create systems that could reason using logic and symbols, falls under this category. The goal is to understand and replicate the processes of thought itself.
Behavior: In contrast, this perspective focuses on the external manifestations of intelligence, i.e., what an AI system does rather than how it thinks. Definitions in this category evaluate AI based on its observable actions and behavior in different situations. The key question is whether the AI system behaves intelligently, effectively, or usefully in its interactions with the environment or in solving tasks. This approach is less concerned with mimicking human thought processes and more focused on achieving intelligent outcomes, regardless of the underlying mechanism. For instance, the Turing Test, which assesses a machine’s ability to exhibit intelligent behavior indistinguishable from that of a human, is a classic example of a behavior-oriented definition of AI. Similarly, modern AI applications in areas like robotics or autonomous driving are often evaluated based on their ability to perform tasks effectively and safely in real-world environments, irrespective of whether their internal processes resemble human cognition.
Orientation Towards Human-like Performance vs. Rationality
Russell and Norvig’s second dimension for classifying AI definitions is the orientation towards human-like performance or rationality.
Human-centered approach: This approach defines AI in relation to human intelligence. It aims to create systems that think and act like humans. The benchmark for success is how closely an AI system can mimic human cognitive abilities and behaviors. Definitions in this category often emphasize capabilities such as learning, problem-solving, creativity, emotional intelligence, and natural language understanding, as these are seen as hallmarks of human intelligence. The goal is to build systems that can perform tasks in a way that would be considered intelligent if done by a human. For example, AI research aimed at developing systems that can engage in natural and fluent conversation with humans, understand human emotions, or exhibit human-like creativity in art or music, aligns with this human-centered approach.
Rationality approach: This approach, in contrast, defines AI in terms of rational behavior. Rationality, in this context, refers to the ability to make decisions and take actions that are expected to achieve the best possible outcome or achieve a specific goal, given the available information and constraints. Definitions focused on rationality emphasize efficiency, optimality, logical reasoning, and problem-solving effectiveness. The benchmark is not necessarily human-like performance but rather optimal performance in achieving defined objectives. AI systems designed to act rationally may or may not resemble human behavior; the primary focus is on achieving the best possible results. For instance, AI systems used in automated trading, resource allocation, or complex planning tasks are often evaluated based on their rationality – their ability to make optimal decisions to maximize profits, efficiency, or goal attainment, even if these decisions are not made in a way that a human would typically approach the problem. The concept of bounded rationality, which acknowledges the limitations of computational resources and information, is also relevant in this context, recognizing that truly optimal rationality may be unattainable in complex real-world scenarios.
Four Categories of AI Definitions Based on the Framework
Russell and Norvig’s framework results in four categories of AI definitions by combining the two dimensions: (1) Thinking vs. Acting and (2) Human-like Performance vs. Rationality.
By crossing these two dimensions – (1) Focus on Thought vs. Action and (2) Orientation towards Human-like Performance vs. Rationality – Russell and Norvig’s framework generates four distinct categories of AI definitions. These categories represent different philosophical and methodological approaches to defining and pursuing Artificial Intelligence. The \(2 \times 2\) matrix formed by these dimensions provides a powerful tool for classifying and comparing various definitions of AI.
Thinking Humanly:
Thinking Humanly: AI systems designed to mimic human thinking processes. The goal is to get the answer right in the same way a human would, often involving cognitive models of human thought.
This category combines the focus on thought processes with the orientation towards human-like performance. AI systems in this category are designed to mimic human thinking processes. The goal is not just to get the answer right, but to get it right in the same way that a human would. This approach often involves developing cognitive models of human thought, exploring areas like cognitive psychology and neuroscience to understand how humans think, solve problems, and make decisions. Early AI efforts to build General Problem Solvers, and more recent work in cognitive architectures that aim to simulate the overall structure of the human mind, fall under this category. A key challenge in this approach is validating whether an AI system truly "thinks like a human," as the internal processes of the human mind are complex and not fully understood.
Thinking Rationally:
Thinking Rationally: AI systems designed to think logically and optimally, regardless of human-like thought processes. The emphasis is on correct reasoning and optimal solutions, rooted in logic and mathematics.
This category combines the focus on thought processes with the orientation towards rationality. AI systems here are designed to think rationally, meaning logically and optimally, regardless of whether these thought processes resemble human thinking. The emphasis is on developing systems that can reason correctly, draw valid inferences, and solve problems logically. This approach is rooted in logic, mathematics, and theoretical computer science. Early AI research in areas like logical reasoning, automated theorem proving, and expert systems, which aimed to encode human expertise in logical rules, exemplify this category. The focus is on the correctness of reasoning and the optimality of solutions, rather than mimicking human thought patterns. However, purely rational approaches can sometimes be limited in dealing with real-world complexity, uncertainty, and the nuances of human-like common sense reasoning.
Acting Humanly:
Acting Humanly: AI systems designed to act in a way indistinguishable from human behavior, exemplified by the Turing Test. Focus is on external behavior rather than internal thought processes.
This category combines the focus on behavior with the orientation towards human-like performance. AI systems in this category are designed to act in a way that is indistinguishable from human behavior. The classic example of this is the Turing Test. An AI system that passes the Turing Test would be considered to be "acting humanly." This approach is less concerned with the internal thought processes and more focused on the external behavior. Areas of AI research like natural language processing (to enable human-like conversation), computer vision (to enable human-like perception), and robotics (to enable human-like physical actions) contribute to building systems that can act humanly. While achieving human-like behavior is a significant goal, some argue that it may not be the most effective path to achieving true intelligence, as it may prioritize superficial mimicry over genuine problem-solving ability.
Acting Rationally:
Acting Rationally: AI systems designed to behave in a way that maximizes goal achievement, given knowledge and beliefs. Rational agents are central, focusing on effectiveness and goal achievement.
This category combines the focus on behavior with the orientation towards rationality. AI systems in this category are designed to act rationally, meaning to behave in a way that maximizes the achievement of their goals, given their knowledge and beliefs. This is the approach that Russell and Norvig advocate for and adopt as the "standard model" in their textbook. Rational agents are central to this perspective. A rational agent is one that acts to achieve the best outcome, or the best expected outcome when there is uncertainty. This approach encompasses a wide range of AI techniques, including machine learning, planning, optimization, and decision theory. Modern AI applications in diverse fields, from autonomous vehicles to recommendation systems to medical diagnosis, often aim to create systems that act rationally to achieve specific objectives. The focus is on effectiveness and goal achievement, and rational action may or may not resemble human behavior. In many cases, acting rationally may involve behaviors that are very different from how humans would act, especially in tasks where humans are limited by cognitive biases, processing speed, or access to information.
These four categories provide a valuable framework for understanding the diverse landscape of AI definitions and approaches. They highlight the different goals and evaluation criteria that can be used when defining and building Artificial Intelligence systems. As we proceed through this course, we will encounter examples of AI research and applications that fall into each of these categories, and we will critically examine the strengths and limitations of each perspective. The lecture concludes here, setting the stage for a more detailed exploration of each of these four definitions in the subsequent session.
Conclusion
This introductory lecture has laid the groundwork for our exploration of Artificial Intelligence. We have covered essential preliminary aspects, ranging from course logistics to the foundational challenge of defining AI itself. Key takeaways from today’s session are summarized as follows:
Course Structure and Instructors: The course is structured into two halves, expertly guided by Professor Stefano Mizzaro and Professor Giuseppe Serra. Professor Mizzaro will cover the theoretical and foundational aspects in the first half, while Professor Serra will delve into applied and advanced topics, including machine learning and reinforcement learning, in the second half. This dual-instructor approach ensures a comprehensive coverage of both theoretical underpinnings and practical applications of AI.
Operational Course Information: We have addressed crucial practical details concerning the course, including the lecture schedule, access to learning resources such as slides and recordings via Microsoft Teams, the role of "Artificial Intelligence: A Modern Approach" as the primary textbook, and the structure and tentative dates for examinations. This logistical overview is designed to equip students with the necessary information for effective participation and preparation throughout the course.
The Complexity of Defining AI: A significant portion of this lecture was dedicated to illustrating the inherent complexities in defining Artificial Intelligence. Through initial student brainstorming and an analysis of media portrayals, we highlighted the diverse and often conflicting perspectives on what AI is and what it should be. This exploration underscores that defining AI is not a straightforward task but a multifaceted challenge that requires careful consideration of various viewpoints.
Russell and Norvig’s Framework for AI Definitions: To navigate the definitional landscape of AI, we introduced Russell and Norvig’s structured framework. This framework categorizes AI definitions along two key dimensions: the focus on ‘thinking’ versus ‘acting’ and the orientation towards ‘human-like performance’ versus ‘rationality’. This framework provides a valuable analytical tool for understanding and comparing different approaches to AI.
Four Categories of AI Definitions: Utilizing Russell and Norvig’s framework, we outlined four distinct categories of AI definitions: Thinking Humanly, Thinking Rationally, Acting Humanly, and Acting Rationally. These categories represent fundamentally different perspectives on the goals and evaluation criteria for Artificial Intelligence. Understanding these categories is essential for appreciating the breadth and depth of the AI field and for critically evaluating different AI approaches and technologies.
Important Remarks and Key Takeaways:
AI as a Vast and Evolving Interdiscipline: It is crucial to recognize Artificial Intelligence as a vast, interdisciplinary, and rapidly evolving field. Its impact spans numerous sectors, and its continuous development necessitates ongoing learning and adaptation. This course aims to provide a robust foundation for navigating this dynamic domain.
Absence of a Singular AI Definition: The lecture emphasized that there is no universally accepted, monolithic definition of Artificial Intelligence. This reflects the field’s inherent complexity, its multifaceted nature, and its continuous evolution. Embracing this diversity of perspectives is key to a nuanced understanding of AI.
"Artificial Intelligence: A Modern Approach" as Guiding Textbook: "Artificial Intelligence: A Modern Approach" by Russell and Norvig will serve as the primary textbook for this course. It will provide a comprehensive and authoritative guide to the theoretical and technical content, structuring our exploration of AI principles and practices.
Follow-up Questions and Topics for the Next Lecture: To prepare for our next session and to deepen your understanding of the topics introduced today, consider the following questions and themes:
In-depth Analysis of the Four AI Definition Categories: We will begin the next lecture with a detailed examination of each of the four categories of AI definitions (Thinking Humanly, Thinking Rationally, Acting Humanly, Acting Rationally). Reflect on the nuances of each category and consider examples of AI systems that might exemplify each approach. What are the inherent strengths and limitations of each perspective?
Prioritization of Definitions in this Course: Which of the four categories of AI definitions outlined by Russell and Norvig will be prioritized in this course, and why? What are the implications of focusing on one definition over others for the course content and learning objectives?
Historical Evolution of AI Definitions and Focus: How have the definitions and primary focus of Artificial Intelligence evolved throughout its history? What were the dominant paradigms in different eras of AI research and development, and how have they shifted over time? Understanding this historical context will provide valuable perspective on the current state of the field.
Ethical and Societal Implications of AI Approaches: What are the ethical and societal implications of pursuing different approaches to AI, particularly concerning the distinction between human-like AI and rationally-acting AI? How do these different approaches shape our expectations, concerns, and regulatory considerations regarding Artificial Intelligence?
Engaging with these questions before the next lecture will significantly enhance your comprehension and participation as we delve deeper into the fascinating and complex world of Artificial Intelligence.