Unit 1: Foundations of Artificial Intelligence - Subjective Questions

Artificial Intelligence (AI) is the branch of computer science concerned with designing systems that can perform tasks requiring human-like intelligence, such as learning, reasoning, perception, language understanding, decision-making, and problem-solving.

Primary goals of AI:

• To develop machines that can perceive and interpret their environment.
• To enable systems to learn from data and experience.
• To solve complex problems efficiently.
• To support or automate decision-making.
• To interact with humans using natural language and other forms of communication.

Characteristics of AI systems:

• Autonomy: Ability to operate with limited human intervention.
• Adaptability: Ability to improve performance based on experience.
• Reasoning: Ability to draw conclusions from available information.
• Perception: Ability to interpret images, speech, sensor data, and other inputs.
• Goal orientation: Ability to select actions that help achieve a desired objective.

AI is significant because it is used in healthcare diagnosis, robotics, fraud detection, recommendation systems, education, transportation, manufacturing, and smart cities. It improves efficiency, supports better decisions, and enables automation of tasks that are difficult or dangerous for humans.

The evolution of AI can be summarized through the following stages:

• Early foundations: AI developed from mathematics, logic, philosophy, neuroscience, and computer science. The development of digital computers made automated reasoning possible.
• Birth of AI as a field: The term Artificial Intelligence was popularized at the Dartmouth conference in 1956. Early researchers focused on symbolic reasoning, theorem proving, and game playing.
• Symbolic AI period: Systems used rules, logic, and manually encoded knowledge. Expert systems were developed for domains such as medical diagnosis and engineering.
• AI winters: Progress slowed during periods when expectations exceeded available computing power, data, and algorithms. Funding and interest declined during these periods.
• Machine learning era: Systems began learning patterns from data rather than depending entirely on manually written rules. Statistical methods, decision trees, support vector machines, and neural networks became important.
• Deep learning era: Powerful processors, large datasets, and improved neural network architectures enabled major advances in computer vision, speech recognition, and natural language processing.
• Generative AI era: Modern models can generate text, images, audio, video, and software code. These systems learn complex patterns from large datasets and produce new content in response to prompts.

The evolution of AI reflects a shift from handcrafted rules to data-driven learning and finally to systems capable of generating complex and varied outputs.

AI can be classified in two common ways: according to its capability and according to its functionality.

Classification based on capability:

• Artificial Narrow Intelligence (ANI): Also called weak AI, it is designed to perform a specific task or a limited set of tasks. Examples include spam filters, virtual assistants, and recommendation systems.
• Artificial General Intelligence (AGI): A theoretical form of AI that would possess human-like intelligence and perform a wide range of intellectual tasks across different domains.
• Artificial Superintelligence (ASI): A hypothetical form of AI that would exceed human intelligence in nearly every area, including creativity, reasoning, and problem-solving.

Classification based on functionality:

• Reactive machines: Respond only to current inputs and do not store past experiences. A basic game-playing system is an example.
• Limited-memory systems: Use historical data for a limited period to make decisions. Most current machine learning and autonomous driving systems belong to this category.
• Theory-of-mind AI: A proposed system that would understand human emotions, intentions, beliefs, and social behavior.
• Self-aware AI: A hypothetical system that would possess consciousness and an understanding of its own internal state.

Most existing AI applications are examples of narrow AI and limited-memory systems. AGI, ASI, theory-of-mind AI, and self-aware AI remain largely theoretical.

These terms are related but represent different scopes and purposes.

• Artificial Intelligence: The broadest field. It aims to create systems capable of intelligent behavior, including reasoning, planning, perception, language understanding, and decision-making.
• Machine Learning (ML): A subfield of AI in which systems learn patterns from data and use them to make predictions or decisions without being explicitly programmed for every rule.
• Deep Learning: A subfield of ML that uses multilayered artificial neural networks. It is especially effective for unstructured data such as images, audio, video, and text.
• Data Science: An interdisciplinary field that uses statistics, programming, domain knowledge, data analysis, visualization, and ML to extract useful insights from data.

Relationship:

Data Science overlaps with AI and ML but is broader in its focus on collecting, cleaning, analyzing, interpreting, and communicating data. For example, a data scientist may use ML to predict customer churn, while an AI engineer may develop an intelligent recommendation system using that prediction model.

An AI problem is commonly formulated as a state-space search problem. The formulation identifies the possible situations, available actions, desired objective, and cost of reaching that objective.

• Initial state: The starting condition of the problem. For example, the current location of a delivery vehicle.
• Actions: The legal operations available from a state, such as moving to a neighboring location.
• Transition model: A description of the state produced after applying an action. It may be deterministic or uncertain.
• Goal test: A procedure that checks whether a state satisfies the required objective.
• Path cost: A numerical measure of the total cost of a sequence of actions, such as distance, time, energy, or money.

A solution is a sequence of actions that transforms the initial state into a goal state. An optimal solution minimizes the path cost. For example, in route planning, the initial state is the starting city, actions are road movements, the goal test checks whether the destination has been reached, and path cost may represent total travel distance.

A well-defined problem formulation reduces unnecessary search and helps an AI system select an effective algorithm.

AI problem-solving techniques search through possible states to find a sequence of actions leading to a goal.

Uninformed search:

• Uses no domain-specific knowledge beyond the problem definition.
• Examples include breadth-first search, depth-first search, uniform-cost search, and depth-limited search.
• Breadth-first search explores the shallowest nodes first and can find the shortest solution when all action costs are equal.
• Depth-first search explores one branch deeply before backtracking and usually requires less memory, but it may not find the best solution.

Informed search:

• Uses additional knowledge, usually represented by a heuristic function, to guide the search.
• Examples include greedy best-first search and A* search.
• A heuristic estimates the cost from node to a goal.
• A* evaluates a node using:
  
  where is the cost from the start to node and is the estimated cost from to the goal.

Uninformed search is simpler but may examine many unnecessary states. Informed search can be more efficient when the heuristic is accurate. A* is complete and optimal under suitable conditions, such as a non-overestimating heuristic.

The AI Development Lifecycle is a structured process for designing, building, deploying, and maintaining AI systems.

1. Problem definition: Identify the business or social problem, objectives, constraints, stakeholders, and success criteria.
2. Data collection: Gather relevant data from databases, sensors, applications, surveys, or public sources while considering consent and privacy.
3. Data preparation: Clean missing or noisy values, remove duplicates, integrate sources, label examples, and transform data into a suitable format.
4. Exploratory data analysis: Study distributions, relationships, outliers, class imbalance, and possible biases.
5. Model selection and development: Select an appropriate algorithm, train the model, and tune its parameters.
6. Evaluation: Test the model using suitable metrics and compare its performance with requirements and baseline methods.
7. Deployment: Integrate the model into an application, service, device, or organizational workflow.
8. Monitoring and maintenance: Track accuracy, fairness, latency, security, and data drift. Retrain or update the model when performance declines.
9. Governance and documentation: Record data sources, assumptions, limitations, decisions, risks, and responsible-use requirements.

The lifecycle is iterative rather than strictly linear. Evaluation or monitoring may reveal the need to return to data preparation, problem definition, or model development.

An intelligent agent is an entity that perceives its environment through sensors and acts upon that environment through actuators to achieve specified goals.

• Sensors: Devices or mechanisms that receive information from the environment. Examples include cameras, microphones, temperature sensors, and keyboard input.
• Percept: The information received by the agent through its sensors at a particular time.
• Percept sequence: The complete history of percepts received by the agent.
• Actuators: Mechanisms used to affect the environment. Examples include motors, robotic arms, speakers, and display systems.
• Actions: Operations selected by the agent and executed through actuators.
• Agent program: The computational process that maps percept sequences to actions.

The general relationship can be represented as:

For example, a robotic vacuum uses sensors to detect obstacles and dirt, processes the sensor information, and uses motors to move or change direction. An agent is considered intelligent when it selects actions that improve its performance with respect to a specified performance measure.

A rational agent is an agent that selects the action expected to maximize its performance measure, given its percept sequence, available knowledge, and possible actions. Rationality does not mean that the agent is perfect or omniscient. It means that the agent makes the best decision possible using the information and resources available.

A performance measure defines what counts as success. For example:

• A medical diagnosis agent may aim to maximize diagnostic accuracy and minimize harmful recommendations.
• A delivery robot may aim to minimize travel time, energy consumption, and package damage.
• A smart traffic system may aim to reduce congestion, waiting time, and emissions.

A rational agent considers:

• The current percept and past percepts.
• Its built-in knowledge and learned information.
• The possible outcomes of available actions.
• The uncertainty of the environment.
• The performance measure used to evaluate results.

The same environment can produce different behavior when the performance measure changes. If a navigation agent is rewarded only for speed, it may choose risky routes. If safety, fuel consumption, and travel time are all included, it is more likely to choose a balanced route. Therefore, designing an appropriate performance measure is essential for safe and useful AI behavior.

An agent environment is the external world in which an agent operates. Its characteristics influence the type of sensors, algorithms, memory, and decision-making methods required.

• Fully observable vs partially observable: In a fully observable environment, the agent receives all information relevant to decision-making. In a partially observable environment, sensors provide incomplete or noisy information.
• Deterministic vs stochastic: In a deterministic environment, an action has a predictable result. In a stochastic environment, outcomes involve uncertainty.
• Episodic vs sequential: In an episodic environment, each decision is independent. In a sequential environment, current actions affect future states.
• Static vs dynamic: A static environment does not change while the agent is deciding. A dynamic environment may change continuously.
• Discrete vs continuous: Discrete environments have a limited number of states or actions, while continuous environments involve values such as speed, time, and position.
• Single-agent vs multi-agent: A single-agent environment has one primary decision-maker. A multi-agent environment includes other cooperative or competitive agents.
• Known vs unknown: In a known environment, the agent understands the rules and outcomes. In an unknown environment, it must learn them.

For example, autonomous driving is partially observable, stochastic, sequential, dynamic, continuous, and multi-agent. Therefore, it requires real-time perception, prediction, planning, and risk management.

AI enables robots to perceive their surroundings, plan actions, learn from experience, and interact with humans.

Applications in robotics:

• Industrial robots perform assembly, welding, painting, and quality inspection.
• Autonomous mobile robots transport materials in warehouses.
• Service robots assist with cleaning, delivery, and customer interaction.
• Medical robots support surgery, rehabilitation, and patient assistance.
• Search-and-rescue robots operate in dangerous environments.
• Agricultural robots monitor crops, remove weeds, and harvest produce.

Benefits:

• Improved precision and productivity.
• Operation in hazardous or repetitive environments.
• Reduced human exposure to danger.
• Continuous operation and consistent performance.
• Support for people with disabilities and elderly users.

Challenges:

• Robots must handle uncertain and changing environments.
• Sensor errors can cause unsafe actions.
• Training and maintenance can be expensive.
• Human-robot interaction requires reliable communication and trust.
• Safety, privacy, employment, and accountability issues must be addressed.

AI-based robots should include testing, fail-safe mechanisms, human supervision when necessary, and clear responsibility for decisions.

AI is increasingly used to support healthcare professionals and improve patient services.

Applications:

• Medical image analysis for detecting tumors, fractures, and other abnormalities.
• Clinical decision support and risk prediction.
• Personalized treatment recommendations.
• Drug discovery and analysis of biological data.
• Remote patient monitoring through wearable devices.
• Chatbots for appointment scheduling and basic health information.
• Robotic assistance in surgery and rehabilitation.

Advantages:

• Faster analysis of large medical datasets.
• Early detection of disease patterns.
• Support for personalized care.
• Reduction of administrative workload.
• Improved access to healthcare services in remote areas.

Limitations and concerns:

• AI may produce incorrect predictions or miss rare conditions.
• Training data may not represent all populations.
• Patient data is highly sensitive and requires strong privacy protection.
• Models may be difficult for doctors and patients to interpret.
• Responsibility must be clear when an AI-supported decision causes harm.
• AI should assist qualified professionals rather than replace clinical judgment in high-risk situations.

Responsible healthcare AI requires validation, explainability, informed consent, security, fairness testing, human oversight, and continuous monitoring.

AI supports manufacturing by analyzing production data, controlling machines, predicting failures, and improving operational decisions.

Major applications:

• Predictive maintenance: AI analyzes vibration, temperature, pressure, and usage data to predict equipment failure before it occurs.
• Quality inspection: Computer vision detects defects, incorrect assembly, and surface damage.
• Process optimization: AI adjusts production parameters to improve quality and reduce waste.
• Demand forecasting: Models estimate future demand and support inventory planning.
• Robotic automation: Intelligent robots perform assembly, packaging, material handling, and inspection.
• Supply-chain optimization: AI improves scheduling, logistics, procurement, and warehouse operations.
• Worker safety: Systems identify unsafe conditions and issue warnings.

Industry 4.0 combines AI with the Internet of Things, cloud computing, robotics, digital twins, and real-time analytics. Sensors continuously collect data from machines, and AI converts this data into predictions or actions.

Benefits include lower downtime, improved product quality, reduced costs, better resource utilization, and greater flexibility. Challenges include cybersecurity threats, integration with older equipment, shortage of skilled personnel, high implementation costs, and the need to manage changes in the workforce responsibly.

A smart city uses digital technologies and data-driven systems to improve the quality, efficiency, sustainability, and safety of urban life.

Applications of AI in smart cities:

• Traffic management: AI analyzes traffic cameras, GPS data, and sensors to optimize signals, detect congestion, and suggest routes.
• Public transportation: Demand prediction and route optimization improve bus and rail services.
• Waste management: Smart bins and predictive scheduling help optimize collection routes.
• Energy management: AI balances electricity demand, integrates renewable energy, and improves building efficiency.
• Water management: Models detect leaks, forecast demand, and support efficient distribution.
• Public safety: Video and sensor analytics can identify hazards, although strict privacy controls are required.
• Environmental monitoring: AI tracks air quality, noise, flooding, and pollution.
• Emergency response: Systems help predict risks and coordinate resources.

AI can reduce costs, emissions, travel time, and resource wastage. However, smart-city systems may create risks involving surveillance, data misuse, cybersecurity, unequal access, and biased decision-making. Cities should use data minimization, transparency, public consultation, strong security, and fairness assessments when deploying AI.

AI ethics is the study and application of moral principles, professional responsibilities, and social values to the design, development, deployment, and use of AI systems.

Important principles include:

• Fairness and non-discrimination: AI should not create unjust disadvantages for individuals or groups. Bias must be identified and reduced.
• Transparency: People should receive understandable information about how and why an AI system is being used.
• Explainability: Where decisions are important, the system should provide meaningful reasons for its outputs.
• Accountability: Organizations and individuals must be responsible for AI outcomes and provide mechanisms for correction.
• Privacy: Personal data should be collected lawfully, used appropriately, and protected from unauthorized access.
• Safety and reliability: Systems should be tested, robust, secure, and able to handle unexpected situations.
• Human autonomy: AI should support human decision-making and should not manipulate or unfairly restrict people.
• Beneficence and social well-being: AI should provide genuine benefits and avoid unnecessary harm.
• Sustainability: Developers should consider energy use, environmental impact, and long-term social effects.

Ethics should be addressed throughout the AI lifecycle rather than treated as an issue only after deployment.

Responsible AI refers to developing and using AI in ways that are lawful, ethical, safe, fair, transparent, secure, and beneficial to society.

Practices for trustworthy AI include:

• Clear purpose and scope: Define the intended use, excluded uses, users, and possible risks.
• Quality data governance: Document data sources, consent, labels, limitations, representativeness, and retention policies.
• Bias assessment: Evaluate performance across relevant demographic and user groups and reduce unjust disparities.
• Privacy protection: Use access controls, encryption, anonymization where appropriate, and data minimization.
• Robust testing: Test normal cases, edge cases, adversarial inputs, failures, and changes in the operating environment.
• Human oversight: Allow qualified people to review, override, or appeal important AI decisions.
• Transparency and documentation: Provide model cards, system documentation, user guidance, and explanations suited to the audience.
• Accountability: Assign responsibilities for development, approval, monitoring, incident response, and correction.
• Continuous monitoring: Check accuracy, fairness, security, drift, and unintended effects after deployment.
• Inclusive design: Involve diverse stakeholders and affected communities in development and evaluation.

Responsible AI is a continuous organizational commitment, not merely a technical feature of a model.

Algorithmic bias occurs when an AI system produces systematically unfair or unjust outcomes for certain individuals or groups. Bias may arise even when developers do not intend to discriminate.

Sources of bias:

• Historical discrimination reflected in training data.
• Underrepresentation of certain populations.
• Incorrect, inconsistent, or subjective labels.
• Sampling bias during data collection.
• Features that act as proxies for sensitive attributes.
• Measurement errors and missing data.
• Objectives or performance measures that ignore fairness.
• Different error costs for different groups.

Effects:

• Unfair hiring, lending, insurance, education, or policing decisions.
• Exclusion of vulnerable communities from services.
• Loss of trust and possible legal consequences.
• Reinforcement of existing social inequalities.

Reduction methods:

• Audit datasets before training.
• Collect more representative and high-quality data.
• Measure performance separately across relevant groups.
• Apply pre-processing, in-processing, or post-processing fairness methods when appropriate.
• Involve domain experts and affected communities.
• Provide human review and appeal mechanisms.
• Monitor deployed systems because bias can emerge as populations or conditions change.

Fairness is context-dependent, so technical metrics must be combined with legal, social, and domain-specific analysis.

Generative AI refers to AI systems that learn patterns from existing data and generate new content that resembles the data on which they were trained. The generated content may be text, images, audio, video, software code, or structured data.

Traditional predictive AI:

• Usually predicts a label, category, value, or probability.
• Examples include classifying an email as spam, predicting house prices, or estimating disease risk.
• Its output is generally a decision, score, or forecast.

Generative AI:

• Produces new content based on a prompt, context, or condition.
• Examples include text-generation assistants, image-generation systems, code-generation tools, music generators, and synthetic-data systems.
• Its output may have many possible valid forms rather than one fixed answer.

Generative models learn a probability distribution over data. In simplified form, they estimate:

Generative AI can support creativity, education, design, customer service, and software development. However, it may produce inaccurate information, reproduce bias, reveal sensitive data, violate intellectual property rights, or be used to create deceptive content. Human review and responsible-use controls are therefore essential.

A large language model (LLM) is a neural-network model trained on a large collection of text to learn relationships between words, symbols, and their contexts.

Basic working principle:

• Text is divided into smaller units called tokens.
• The model processes token sequences and learns contextual relationships.
• During training, it commonly learns to predict the next token based on previous tokens.
• Repeated training enables the model to represent grammar, facts, styles, and patterns in language.
• During generation, the model selects probable next tokens until an answer is formed.

Role of training data:

• Training data determines the knowledge, language patterns, biases, and limitations learned by the model.
• High-quality, diverse, legally obtained, and appropriately filtered data improves reliability.

Role of prompts:

• A prompt provides instructions, context, examples, constraints, or a desired format.
• Clear prompts can improve relevance and reduce ambiguity, but they cannot guarantee factual correctness.

Role of human feedback:

• Human feedback can be used to improve helpfulness, safety, instruction-following, and refusal of harmful requests.
• Reviewers may compare outputs or identify undesirable behavior.

LLMs do not automatically verify every statement. They can generate fluent but incorrect content, so important outputs require verification and human judgment.

Rule-based AI systems use explicitly written rules, logic, facts, and inference procedures. Machine-learning-based systems learn patterns or decision boundaries from examples and data.

Advantages of rule-based AI:

• Predictable behavior.
• Straightforward explanations.
• Useful when regulations and domain rules are clearly defined.

Limitations of rule-based AI:

• Difficult to maintain when rules become numerous.
• Performs poorly with ambiguity, noise, and unanticipated situations.
• Does not automatically learn from experience.

Advantages of machine learning:

• Can identify complex patterns.
• Handles large-scale data and noisy inputs.
• Can improve when new representative data becomes available.

Limitations of machine learning:

• Requires data, validation, and monitoring.
• May inherit bias or produce opaque decisions.
• Can fail when real-world data differs from training data.

Many practical systems combine rules with machine learning to achieve both flexibility and control.