Artificial Intelligence, Machine Learning and Deep Learning

Introduction and Context

Artificial intelligence (AI), machine learning (ML), and deep learning (DL) are closely related but distinct fields. They are often used interchangeably in popular media, yet each term describes a different scope of ideas, methods, and capabilities. Understanding their relationship is essential for anyone working with intelligent digital systems.

AI is the broadest concept. It concerns the design of systems that perform tasks normally associated with human intelligence, such as perception, language use, reasoning, planning, learning, and decision-making. Machine learning is a major approach to AI in which systems learn patterns from data rather than being programmed with every rule explicitly. Deep learning is a specialised branch of machine learning that uses multi-layer artificial neural networks to learn increasingly complex representations from data.

These technologies support digital transformation in sectors such as healthcare, banking, manufacturing, retail, education, logistics, agriculture, and public administration. However, their practical value depends not only on technical performance but also on data quality, organisational readiness, security, explainability, cost, regulation, and responsible governance.

Learning Objectives

After studying this chapter, you should be able to:

1. Distinguish artificial intelligence, machine learning, and deep learning.
2. Describe major stages in the historical development of intelligent systems.
3. Explain important terminology used in AI and neural networks.
4. Compare rule-based, machine-learning, and deep-learning approaches.
5. Identify common capabilities and limitations of AI systems.
6. Explain how AI contributes to digital transformation.
7. Discuss practical, ethical, legal, and governance considerations.
8. Evaluate whether an AI, ML, or DL solution is appropriate for a business problem.

The Conceptual Relationship Among AI, ML and DL

The relationship can be represented as nested categories:

• Artificial intelligence is the broad field of creating systems capable of intelligent behaviour.
• Machine learning is a subset of AI based on learning from examples or experience.
• Deep learning is a subset of machine learning based primarily on neural networks with multiple computational layers.

Thus, every deep-learning system is a machine-learning system, and every machine-learning system is generally considered an AI system. The reverse statements are not true. A chess program based entirely on manually designed search rules may be AI without using machine learning. A simple linear regression model may be machine learning without being deep learning.

Figure: Deep learning is contained within machine learning, which is contained within artificial intelligence.

The boundaries are conceptual rather than absolute. For example, modern systems may combine symbolic rules, statistical learning, neural networks, optimisation, search, and human feedback. Such systems are often called hybrid AI or neuro-symbolic AI.

What Is Artificial Intelligence?

AI is the study and engineering of computational systems that exhibit capabilities commonly associated with human intelligence. These capabilities include:

• Perception: interpreting images, sounds, sensor readings, or text.
• Knowledge representation: storing facts, concepts, relationships, and rules.
• Reasoning: drawing conclusions from available information.
• Planning: selecting actions to achieve a goal.
• Language processing: understanding or generating human language.
• Learning: improving performance from data or experience.
• Decision-making: choosing among alternatives under conditions of uncertainty.

The term “intelligence” does not imply consciousness, emotion, or human-like understanding. An AI system may perform a narrow task extremely well while lacking general reasoning ability. A system that detects fraudulent transactions, for instance, may not be able to explain a medical image or control a robot.

Narrow and General AI

Narrow AI, also called weak AI, is designed for a defined task or a limited group of related tasks. Examples include recommendation engines, route-planning software, speech recognition, and automated quality inspection.

Artificial general intelligence (AGI) refers to a hypothetical system with broad, flexible intellectual capabilities comparable to or exceeding those of humans across many domains. Current commercial systems should not automatically be described as AGI. High performance in one domain is not evidence of general intelligence.

Symbolic and Data-Driven AI

Early AI commonly used symbolic methods. In a symbolic system, knowledge is represented explicitly using rules, symbols, logic, or structured relationships. For example:

• IF a machine temperature exceeds a threshold, THEN issue an alert.
• IF a customer has missed several payments, THEN classify the account as high risk.

Symbolic systems can be transparent and effective when rules are stable and well understood. They become difficult to maintain when situations are numerous, uncertain, or constantly changing.

Data-driven AI learns statistical relationships from examples. It is particularly useful where explicit rules are difficult to write, such as recognising objects in images or predicting demand. Its main risks include biased data, limited explainability, and poor performance outside the conditions represented in training data.

What Is Machine Learning?

Machine learning is a computational approach in which a model learns a mapping, structure, or decision policy from data. Instead of specifying every decision rule, a developer selects a learning method, supplies appropriate data, defines an objective, and evaluates the resulting model.

A useful abstraction is:

Here, x is an input, ŷ is the model’s prediction, fθ is a function controlled by parameters θ, and ŷ is an estimate of the desired output. During training, the parameters are adjusted so that predictions become more useful according to a chosen objective. The notation does not mean that learning is automatic or error-free; the result depends strongly on data, features, model assumptions, and evaluation procedures.

Main Learning Paradigms

#### Supervised learning

In supervised learning, the training data contain examples of inputs and known target outputs. The model learns to predict the target for new inputs.

Examples include:

• predicting house prices from property characteristics;
• classifying an email as spam or legitimate;
• estimating whether a component will fail within a specified period;
• identifying a disease category from labelled medical images.

For a regression problem, the target is numerical. For a classification problem, the target is a category. Binary classification has two classes, while multiclass classification has more than two.

#### Unsupervised learning

Unsupervised learning uses data without labelled target outputs. The system seeks structure within the data, such as clusters, dimensions, or unusual observations.

Common applications include customer segmentation, anomaly detection, document grouping, and data visualisation. The discovered patterns require human interpretation; a cluster is not automatically a meaningful business segment.

#### Semi-supervised learning

Semi-supervised learning combines a small amount of labelled data with a larger amount of unlabelled data. It is useful when labelling requires expensive expert effort, as in medical diagnosis or industrial inspection.

#### Self-supervised learning

In self-supervised learning, labels are generated from the data itself. For example, part of a sentence may be hidden and the model learns to predict the missing content. This approach has become central to training large language and multimodal models.

#### Reinforcement learning

Reinforcement learning concerns an agent that interacts with an environment. The agent observes a state, takes an action, receives a reward or penalty, and seeks a policy that produces high long-term reward. Applications include robotics, game-playing, resource allocation, and adaptive control. Reinforcement learning is different from supervised learning because the correct action is not necessarily provided for every situation.

Important Machine-Learning Terminology

Dataset, observation, feature and label

A dataset is a collection of examples. An individual example is an observation, record, or instance. A feature is an input variable used by a model. A label, target, or response is the desired output in supervised learning.

For a loan-risk system, monthly income, existing debt, and repayment history may be features, while the observed repayment outcome may be the label.

Model, parameter and hyperparameter

A model is a mathematical or computational representation used to produce predictions or decisions. A parameter is learned from data, such as a weight in a neural network. A hyperparameter is selected before or during training, such as learning rate, tree depth, or number of neural-network layers.

Training, validation and test data

The training set is used to estimate model parameters. The validation set is used to compare alternatives and tune hyperparameters. The test set is reserved for final evaluation. The test set should not influence model selection; otherwise, the reported performance may be overly optimistic.

Generalisation and overfitting

Generalisation is the ability to perform well on new data from the intended operating environment. Overfitting occurs when a model learns details or noise specific to the training data and consequently performs poorly on new cases. Underfitting occurs when the model is too simple to capture relevant structure.

A common conceptual relationship is:

This is a simplified bias–variance decomposition under particular squared-error assumptions. “Irreducible noise” represents variation that cannot be predicted from the available inputs, bias² represents systematic error from restrictive assumptions, and variance represents sensitivity to the particular training sample. The expression is a conceptual guide rather than a universal identity for every metric or learning problem.

Data leakage and distribution shift

Data leakage occurs when information unavailable at prediction time enters training or evaluation. For example, using a hospital discharge code to predict whether a patient will be admitted leaks future information into the model.

Distribution shift occurs when the statistical properties of operational data differ from those of the training data. Changes in customer behaviour, sensor calibration, economic conditions, or regulations can reduce accuracy. Monitoring is therefore required after deployment.

What Is Deep Learning?

Deep learning uses artificial neural networks with multiple learned layers. Each layer transforms its input into a new representation. Early layers may detect simple patterns, while later layers combine them into more complex structures. In image recognition, early computations may respond to edges, intermediate layers to shapes, and later layers to objects. This description is intuitive rather than a guarantee that each layer has a simple human-interpretable meaning.

A basic neuron computes a weighted sum followed by an activation function:

Here, x₁, x₂, …, xₙ are input values, w₁, w₂, …, wₙ are learned weights, b is a bias term, z is the pre-activation value, φ is an activation function, and a is the neuron output. The variables must be represented on compatible numerical scales; normalisation is often useful, although the appropriate preprocessing depends on the data and model.

A network is trained by comparing predictions with target values using a loss function. For a dataset of m examples, a general empirical loss can be written as:

Here, J(θ) is the average training loss, m is the number of examples, xᵢ is an input, yᵢ is its target, fθ(xᵢ) is the prediction, and L measures prediction error. Training seeks parameters that reduce this objective, usually through an optimisation method such as gradient descent. The formula applies to an average loss over a finite dataset; it does not itself guarantee good generalisation.

Common Deep-Learning Architectures

• Feedforward networks: information moves from input to output without recurrent connections. They are common for tabular prediction and basic classification.
• Convolutional neural networks: use local connectivity and shared filters, making them effective for images and spatial data.
• Recurrent neural networks: process sequences while maintaining an internal state. They have been used for speech, time series, and language, although transformers are now dominant in many language applications.
• Transformers: use attention mechanisms to model relationships among elements in a sequence or set. They support large language models, translation, document analysis, and multimodal systems.
• Autoencoders: learn to encode data into a lower-dimensional representation and reconstruct it. They may support compression, representation learning, or anomaly detection.
• Generative models: produce new text, images, audio, video, or structured data based on learned distributions.

Deep learning is powerful because it can learn representations directly from relatively unprocessed data and can exploit large datasets and specialised hardware. It is not automatically superior. For small structured datasets, a simpler model may be cheaper, more explainable, and equally accurate.

Evolution of AI and Neural Networks

The history of AI is not a simple progression from failure to success. It consists of changing methods, expectations, resources, and application contexts.

1. Foundational ideas: Mathematical logic, statistics, cybernetics, information theory, and early theories of computation supplied intellectual foundations.
2. Symbolic AI: From the 1950s, researchers developed programs for theorem proving, game playing, planning, and problem solving using rules and search.
3. Early neural networks: Perceptron models demonstrated that machines could learn simple decision boundaries, but early systems had important limitations.
4. Expert systems: In the 1970s and 1980s, rule-based systems encoded specialist knowledge for domains such as diagnosis and configuration. Their maintenance burden contributed to periods of reduced enthusiasm known as AI winters.
5. Statistical machine learning: From the 1990s, improved data availability and algorithms such as decision trees, support-vector machines, probabilistic models, and ensemble methods broadened practical applications.
6. Modern deep learning: Large datasets, graphics processing units, improved optimisation, better architectures, and open-source software produced major advances in speech, vision, language, and generation.
7. Foundation and multimodal models: Contemporary systems are often pre-trained on broad data and adapted to many tasks. Their scale increases capability but also raises concerns about cost, documentation, copyright, privacy, misuse, and evaluation.

The history shows that AI progress depends on more than algorithms. It also depends on computing infrastructure, suitable data, research methods, investment, social acceptance, and clear problem definition.

Capabilities and Limitations

AI systems can classify, predict, recommend, search, generate content, optimise schedules, detect anomalies, and support human decisions. Their strengths often include speed, consistency, scalability, and the ability to process large volumes of information.

However, several limitations must be recognised:

• Dependence on data: Poor, incomplete, biased, or outdated data produce unreliable results.
• Limited causal understanding: A model may identify association without establishing that one factor causes another.
• Context sensitivity: Performance may decline in unfamiliar environments or unusual cases.
• Hallucination and fabrication: Generative systems may produce fluent but false statements.
• Opacity: Complex models may be difficult to interpret at the level required by users or regulators.
• Security vulnerabilities: Adversaries may manipulate inputs, steal models, poison training data, or exploit poorly protected interfaces.
• Resource demands: Training and operating large models can require substantial electricity, hardware, and specialist expertise.
• Automation bias: People may accept a system’s recommendation too readily, even when it is wrong.
• Unequal effects: Errors may be distributed unevenly among demographic or socioeconomic groups.

A high accuracy score alone is insufficient. Evaluation should consider error types, subgroup performance, calibration, robustness, latency, cost, usability, and the consequences of failure.

AI in Digital Transformation

Digital transformation is the organisational redesign of processes, services, capabilities, and business models through digital technologies. AI contributes to this transformation in several ways:

1. Process automation: repetitive document processing, invoice matching, customer support triage, and visual inspection.
2. Decision support: demand forecasting, credit assessment, predictive maintenance, and clinical prioritisation.
3. Personalisation: product recommendations, adaptive learning, and targeted communication.
4. Prediction and optimisation: route planning, inventory control, energy management, and workforce scheduling.
5. New products and services: intelligent assistants, machine-generated content, language translation, and personalised health applications.
6. Knowledge access: search, summarisation, question answering, and enterprise knowledge management.

Successful transformation is socio-technical. An organisation must redesign workflows, define accountability, train employees, integrate data systems, protect information, and measure business outcomes. Installing a model without changing the surrounding process may create little value or may introduce new risk.

A useful implementation sequence is:

Figure: AI transformation is a lifecycle rather than a one-time software installation.

Ethical, Legal and Governance Considerations

Responsible AI requires attention throughout the system lifecycle.

Fairness: Training data and decision criteria should be assessed for unjustified disparities. Fairness is context-dependent; a single statistical measure cannot resolve every ethical question.

Transparency and explainability: Users should know when AI is being used, what it is intended to do, and what its important limitations are. High-stakes decisions may require explanations that are meaningful to affected people.

Privacy: Organisations should collect only necessary data, establish lawful purposes, control access, protect stored information, and consider anonymisation or privacy-preserving techniques.

Accountability: A human or organisation must remain responsible for outcomes. “The algorithm decided” is not an adequate governance arrangement.

Safety and security: Systems should be tested against foreseeable misuse, adversarial inputs, failures, and changing conditions. Access controls, logging, incident response, and model monitoring are essential.

Intellectual property and provenance: Training and generated content may raise copyright, licensing, attribution, and ownership questions. Organisations should document data sources and usage rights.

Human oversight: Oversight must be substantive rather than symbolic. Users need authority, competence, time, and procedures to challenge or override automated recommendations.

Regulatory requirements differ by jurisdiction and sector. Therefore, practitioners should consult applicable data-protection, consumer-protection, employment, safety, medical-device, financial, and AI-specific rules rather than assuming that a technically successful model is legally deployable.

Illustrative Case: Predictive Maintenance in a Manufacturing Plant

Background

A manufacturing plant operates hundreds of electric motors. Unexpected failures stop production and require expensive emergency repairs. The plant records vibration, temperature, operating speed, maintenance history, and failure events.

Problem

Managers want to identify motors at elevated risk of failure during the next seven days. The objective is not merely to maximise accuracy; false negatives may cause downtime, while excessive false positives may lead to unnecessary maintenance.

Analysis

A data scientist first checks whether sensor readings are time-stamped consistently and whether maintenance records are complete. The team defines the prediction time carefully. Measurements taken after a failure must not be used to predict that same failure, because doing so would create leakage.

The data are divided chronologically rather than randomly. A chronological split better represents deployment, where future failures must be predicted from past information. Candidate models include logistic regression, a tree-based model, and a neural network for sequential sensor data.

Approach and Solution

The team begins with a simple baseline: an alert based on a temperature threshold. It then develops supervised models using historical windows of sensor measurements. Evaluation includes recall for actual failures, false-alert rate, lead time, performance by motor type, and maintenance cost.

Suppose the operational policy prioritises avoiding missed failures. The model threshold is adjusted to increase recall, but alerts are routed to maintenance engineers rather than triggering automatic shutdown. Engineers review sensor trends and physical conditions before acting. The system also records whether an alert was accepted, rejected, or incorrect.

Outcome

In this illustrative scenario, the model identifies many failures earlier than the threshold rule, but performance declines on a newly installed motor type. The plant therefore adds a monitoring rule for distribution shift and requires a review before extending the model to new equipment. The system becomes a decision-support tool, not an unquestioned replacement for engineering judgement.

Learning Points

1. The business objective must be translated into measurable costs and benefits.
2. Time-aware data splitting prevents misleading evaluation.
3. A baseline is necessary for judging whether a complex model adds value.
4. Human review can reduce operational risk in high-consequence settings.
5. Deployment requires monitoring, feedback, documentation, and retraining policies.

Best Practices for Selecting an AI Approach

1. Define the decision, user, time horizon, and consequence of error.
2. Establish a non-AI baseline before selecting a sophisticated model.
3. Audit data quality, provenance, representativeness, and legal permissions.
4. Use training, validation, and genuinely independent test data.
5. Select metrics that reflect the real cost of errors.
6. Prefer the simplest model that meets the required performance and governance needs.
7. Test on realistic edge cases and relevant demographic or operational groups.
8. Document assumptions, limitations, version changes, and responsible owners.
9. Design human oversight and escalation procedures before deployment.
10. Monitor performance, drift, security, fairness, cost, and user feedback after deployment.

Contemporary Developments

Current AI development is shaped by foundation models, generative AI, multimodal systems, retrieval-augmented generation, edge AI, federated learning, and responsible-AI engineering. Foundation models are trained broadly and adapted to many tasks. Retrieval-augmented systems combine generation with access to an external knowledge base, which can improve factual grounding but does not eliminate errors. Edge AI places inference near sensors or users, reducing latency and sometimes improving privacy.

The field is also moving toward stronger evaluation, model documentation, data governance, energy-efficiency research, and specialised smaller models. These developments reinforce a central principle: capability must be considered together with reliability, safety, cost, legality, and social impact.

Chapter Summary

Artificial intelligence is the broad field of systems that perform tasks associated with human intelligence. Machine learning is an AI approach that learns patterns from data, while deep learning is machine learning based mainly on multi-layer neural networks. AI includes symbolic, statistical, hybrid, and learning-based methods.

Machine learning may be supervised, unsupervised, semi-supervised, self-supervised, or reinforcement-based. Important concepts include features, labels, parameters, hyperparameters, training and test sets, generalisation, overfitting, data leakage, and distribution shift. Deep learning learns layered representations and is especially effective with large-scale unstructured data, although it can be expensive, opaque, and data-hungry.

AI has evolved through symbolic reasoning, expert systems, statistical learning, and modern deep learning. It supports digital transformation through automation, prediction, personalisation, optimisation, and new services. Responsible deployment requires attention to privacy, fairness, security, explainability, accountability, human oversight, legal compliance, and ongoing monitoring. The best solution is not always the most complex model; it is the approach that solves the defined problem reliably and responsibly.

Key Terms

Artificial intelligence: The field of creating systems capable of tasks associated with human intelligence.

Machine learning: An approach in which models learn patterns or decisions from data.

Deep learning: Machine learning using neural networks with multiple learned layers.

Narrow AI: AI designed for a limited task or domain.

Artificial general intelligence: A hypothetical system with broad, flexible intelligence across domains.

Symbolic AI: AI based on explicit symbols, rules, logic, or structured knowledge.

Supervised learning: Learning from examples containing inputs and known targets.

Unsupervised learning: Learning structure from data without target labels.

Reinforcement learning: Learning actions through interaction, rewards, and penalties.

Feature: An input variable used by a model.

Label: The desired target output in supervised learning.

Parameter: A value learned by a model during training.

Hyperparameter: A configuration value selected outside the learning process.

Overfitting: Learning training-specific detail that harms performance on new data.

Generalisation: Performance on new data from the intended environment.

Data leakage: Unintended use of information unavailable at prediction time.

Distribution shift: Change between training-data and operational-data distributions.

Neural network: A layered computational model composed of interconnected units.

Foundation model: A broadly pretrained model adaptable to multiple tasks.

Digital transformation: Organisational change enabled by digital technologies and redesigned processes.

Review Questions

Short Answer Questions

1. Define artificial intelligence, machine learning, and deep learning, and state their relationship.
2. What is the difference between supervised and unsupervised learning?
3. Explain overfitting in your own words.
4. What is data leakage, and why does it produce misleading evaluation?
5. State two benefits and two limitations of deep learning.

Long Answer Questions

1. Explain the historical evolution from symbolic AI to modern deep learning.
2. Compare symbolic AI, conventional machine learning, and deep learning in terms of data, interpretability, and applications.
3. Describe the role of training, validation, and test datasets in model development.
4. Explain how AI can contribute to digital transformation and why organisational change is necessary.
5. Discuss the ethical, legal, and governance requirements for deploying AI in a high-stakes domain.

Analytical / Application Questions

1. A college wants to predict which students may require academic support. Propose an AI approach, identify suitable features and risks, and explain how you would evaluate it fairly.
2. A retailer has only 500 labelled images but 100,000 unlabelled product images. Which learning strategies could be considered, and what additional checks would be required before deployment?
3. A factory reports excellent test accuracy for a predictive-maintenance model, but performance falls sharply after new machines are installed. Analyse possible causes and recommend a monitoring and improvement plan.

Multiple Choice Questions with Answers

Question 1

Which statement best describes the relationship among AI, machine learning, and deep learning?

A. AI is a subset of deep learning.

B. Deep learning is a subset of machine learning, which is a subset of AI.

C. Machine learning and AI are unrelated fields.

D. Deep learning includes all symbolic AI systems.

Correct Answer: B. Deep learning is a subset of machine learning, which is a subset of AI.

Question 2

Which learning paradigm uses examples with known target outputs?

A. Supervised learning

B. Unsupervised learning

C. Reinforcement learning only

D. Clustering without labels

Correct Answer: A. Supervised learning

Question 3

What is the main purpose of a validation dataset?

A. To replace the training data

B. To tune and compare model alternatives

C. To guarantee that data contain no bias

D. To provide future operational data

Correct Answer: B. To tune and compare model alternatives

Question 4

A model performs very well on training examples but poorly on new examples. What is the most likely problem?

A. Underfitting

B. Overfitting

C. Data compression

D. Normalisation

Correct Answer: B. Overfitting

Question 5

Which situation is an example of data leakage?

A. Using past sensor readings to predict future failure

B. Removing duplicate records before training

C. Using a discharge code to predict a hospital admission that occurred earlier

D. Separating data into training and test sets

Correct Answer: C. Using a discharge code to predict a hospital admission that occurred earlier

Question 6

What does a parameter in a neural network generally represent?

A. A value learned during training

B. A legal requirement for deployment

C. A user’s password

D. The number of records in a database

Correct Answer: A. A value learned during training

Question 7

Which characteristic is especially associated with deep learning?

A. It requires all rules to be written manually.

B. It uses multi-layer neural networks to learn representations.

C. It can operate without data or computation.

D. It is limited to accounting applications.

Correct Answer: B. It uses multi-layer neural networks to learn representations.

Question 8

Why might a simple model be preferred over a deep neural network for some business problems?

A. Simple models never make errors.

B. Simple models may be cheaper and easier to explain while meeting performance needs.

C. Deep networks cannot process numerical data.

D. Simple models do not require evaluation.

Correct Answer: B. Simple models may be cheaper and easier to explain while meeting performance needs.

Question 9

Which practice best supports responsible AI deployment?

A. Deploying immediately after training accuracy is measured

B. Removing human oversight from high-stakes decisions

C. Monitoring performance, drift, security, and subgroup outcomes after deployment

D. Using the most complex available model regardless of need

Correct Answer: C. Monitoring performance, drift, security, and subgroup outcomes after deployment

Question 10

In reinforcement learning, an agent primarily learns through:

A. Known labels for every possible action

B. Interaction with an environment and received rewards or penalties

C. A fixed set of handwritten image features only

D. Randomly deleting training examples

Correct Answer: B. Interaction with an environment and received rewards or penalties