Philosophy and Ethics of AI Quiz

The Philosophy and Ethics of Artificial Intelligence

• This course covers the philosophical and ethical aspects of Artificial Intelligence (AI)
• The course will cover topics such as consciousness, reasoning, communication, ethics, morality, algorithms, applications in medicine and war, superintelligence, and the value loading problem.
• The course will also cover AI and human society, and review/revision.

Overview of Course

• The course will cover Introduction to AI
• The course will cover the concepts of intelligence and consciousness
• The course will cover Reasoning and Communication
• The course will cover Ethics and Morality
• The course will cover Algorithms and their implications
• The course will cover situations where AI is used for good or bad, like AI in Medicine, AI in War and other applications.
• The course will contain the topic of Superintelligence and the value loading problem.
• The course will cover the topic of AI and Human Society and will also contain a review/revision session.

Reasoning and Communication Paradigms

• Course will contain the major AI paradigms
• The course will cover the limitations of AI paradigms.
• The course will cover argumentation and communication as a form of artificial intelligence.
• The course will cover topics around argumentation and epistemology (This is not part of the assessed material)

Sources

• The course will reference the Stanford Encyclopedia of Philosophy
• P. M. Dung's work on acceptability of arguments, logic programming and n-person games in Artificial Intelligence
• S. Modgil's work on dialogical scaffolding for human and artificial agents in Artificial Intelligence and Cognition.

Logical (Symbolic) Approaches to AI

• Early AI logic-based systems involved using symbolic representations of beliefs, desires, goals and states of affairs.
• Using logical language and inference rules is analogous to natural language.
• The use of logic in mathematics will be examined.

Logic and the Formalisation of Mathematics

• There were mathematical paradoxes at the beginning of the 20th century.
• Research to prove truths in mathematics is from given sets of axioms.
• Systems must be complete (provable) and consistent.
• Gödel's incompleteness theorems showed that this was impossible.
• Axioms such as ∝∨¬∝ and inference rules such as α∨β|α are used to deduct complex statements.
• Euclid proved truths of Euclidean geometry using 5 basic axioms.

Logic and the Formalisation of Mathematics (Example)

• A triangle is used to show the use of logic in mathematics.
• In the example the set of edges and corners of a triangle are symbolically represented.
• Axiomatic system with basic truths (axioms) are given to create a symbolic representation.
• Rules of logical inference are used to prove a conclusion.

From Logic in Mathematics to Logic in AI

• Proof system and model (triangle) are defined.
• If a deduced statement is true about the model, the proof system is sound.
• If all true statements about the model can be inferred, the proof system is complete.
• Good old fashioned AI (GOFAI) used a similar methodology to mathematics.
• Given a model, encode axiomatic truths, which should be consistent with beliefs, and design a proof system to derive further truths.

Logic in AI

• Given an agent's goals and beliefs, the agent acts resulting in a change in the world
• Then new facts are obtained and the knowledge base is updated
• Agents can act and sense the world to continue this process.
• How can logic achieve all of cognition?
• Turing computation can be formalized using predicate logic.
• Intelligence could be formalized using logic.

Monotonic and Non-Monotonic Logics

• Logic in mathematics describes static models/worlds
• In AI, models can change
• Classical Logic is monotonic.
• Real-world beliefs involve exceptions.
• Rules of thumb can be used and withdrawn when exceptions arise
• Non-monotonic logics are required to handle this.
• Non-monotonic logics include defeasible rules, such as 'birds usually fly' as a default rule.

Non-monotonic Logics

• AI uses non-monotonic logics to handle exceptions in rules.
• Non-monotonic logics use defeasible rules of the form "birds typically fly".
• The idea is that birds normally fly, unless there is evidence to the contrary.

Model Theory/Semantics for Logics

• Model theory defines interpretations for first-order/predicate classic logic.
• The interpretation includes a domain of individuals (constants) in the language.
• N-ary tuples of individuals determine truth of n-ary predicates..
• Mappings from sets of individuals define n-ary functions.

Modal Logics

• Modal logics expand propositional and predicate logic with modality operators.
• A modality qualifies a statement, such as 'Sanjay is usually happy'
• Alethic modalities deal with truth, including possibility and necessity.
• Deontic modalities involve obligation, permission, and prohibition, relating to ethics.

Beliefs, Desires and Intentions (BDI)

• BDI logics describe agents' mental attitudes.
• Based on philosopher Michael Bratman's work on practical reasoning.
• Beliefs are about the world, including the agent itself.
• Desires are motivational states (e.g., to become rich).
• Goals are actively pursued desires (must be consistent).
• Intentions are plans to achieve goals.

Machine Learning Approaches to AI

• Logic-based initially did not yield impressive results
• Recent developments incorporate new algorithms, networks, and hardware.
• Google DeepMind's AI systems like AlphaZero have achievements.
• These include translation, language recognition, autonomous vehicles, computer vision, medical image interpretations, text generation, and deep learning robots.
• A new era of success in AI is happening, with notable uses being evident.

Machine Learning- A Very Brief Review of Fundamentals

• Programs can learn from experiences, improving task performance over time.
• Supervised Learning uses input and output data to build a model.
• Semi-Supervised Learning uses incomplete or partially labelled data.
• Unsupervised Learning finds structure or patterns in the data.
• Reinforcement Learning learns from feedback (rewards/penalties) in a dynamic environment.

Limitations of Logic-based GOFAI

• Requiring many rules to be encoded for tasks is problematic.
• Systems are brittle, and a small error leads to significant problems.
• Representation and grounding have problems as rules and data must be explicitly coded.

Limitations of Machine Learning

• Machine learning needs a massive dataset.
• Learning from small datasets should be considered
• Generalization/transfer is a problem; trained networks can perform poorly on related tasks.
• High-level cognition, such as planning, reasoning, and analogy, remains a struggle.

Limitations of Machine Learning (cont.)

• Black box problem: It's hard to understand why and how machine learning methods reach a conclusion.
• Lack of symbolic/language-like explanations makes it hard for machine and human reasoning to be combined.
• Systems struggle with common sense reasoning, which is a fundamental challenge in AI.

Problem of Commonsense Reasoning

• Commonsense reasoning is a fundamental problem for machine learning and logic-based AI.
• It is complex and difficult for systems to understand reasoning effectively.
• Non-monotonic logics can address these issues theoretically but are complex and computationally demanding.

Argumentation and Communication

• A new approach to non-monotonic reasoning, argumentation.
• Argumentation is more human-like than previous approaches.
• It can handle computational limitations in formalizing/executing reasoning tasks.
• Integrating human and machine reasoning is done via dialogues.
• This is important for understanding how AI can be ethical.

Non-monotonic Reasoning as Argumentation

• Non-monotonic reasoning is about dealing with conflicting beliefs, desires or goals.
• Choosing rationally amongst these is a core component of intelligence.
• Argumentation and debates are examples of non-monotonic reasoning.
• Proofs are analyzed as arguments with grounds/premises that support the claim.
• A1 and A2 are examples of arguments, which can be 'attacked'.

Dung's theory of Argumentation

• Formalizes non-monotonic reasoning as argumentation.
• Define arguments and how they are attacked/supported to create an Argument Framework.
• Rules are used to define a legal labelling that is valid.
• A grounded/preferred labelling determines the winning arguments.

Argument Game Proof Theories

• This is a way to implement Dung's theory.
• The PRO and OPP play a game to evaluate an argument.
• The winner determines whether an argument is in the extension.

Single Agent Non-monotonic Reasoning as Argumentation

• Non-monotonic reasoning in a single agent is described as a proof theory based on an argument framework constructed from the KB.
• Argumentation is used to determine if a claim is valid, which can be part of a larger inference process that could involve multiple agents.

From Single Agent Reasoning to Distributed Reasoning via Dialogue

• Adaptation of argument games is employed for developing dialogical models of distributed reasoning.
• The 'public' aspects of communications and reasoning are defined.
• So that the claim to be demonstrated will be implied from the contents communicated during the dialogue, not just the private knowledge of an agent involved in the dialogue.

Argumentation and Communication (cont.)

• This area of research offers a way to improve the reasoning of AI and human understanding.
• It also offers a way to make the reasoning processes more public and visible.
• This process can provide models for ethical reasoning that are more robust than current models, in particular within AI.

This quiz explores key concepts and limitations related to deep learning and artificial intelligence, focusing on philosophical and ethical considerations. Test your knowledge on topics such as the black box problem, commonsense reasoning, and cognitive processes in AI. Delve into the foundational issues that shape our understanding of machine learning and its implications.