Deep Learning, Generative AI and Intelligent Applications
Free to read online. For a copy of this book, or to quote, highlight or reproduce any part of it, please request permission through the contact page beforehand.
Editorial and publication note
This is a proposed educational series. IEEE-style technical writing and citations do not imply IEEE endorsement, publication, accreditation, certification, or affiliation. Any standard, designation, author credential, publication detail, or institutional relationship must be verified before publication. These Word files are expanded educational drafts and chapter-development manuscripts for editorial review; they are not represented as final peer-reviewed or publication-ready books.
The proposed series title is descriptive only. Do not use IEEE logos or imply IEEE affiliation without authorisation. Any references to standards should be checked in the official IEEE Standards Association catalogue and used only within their stated scope.
About this book
Explores neural networks, optimisation, CNNs, attention, transformers, language models, retrieval, fine-tuning, evaluation, multimodal systems, and deployment.
Who should use this book
This volume is designed for advanced undergraduate and postgraduate learners. Instructors should adapt pacing, depth, tools, and examples to learner needs and local curriculum requirements.
Learning outcomes
- Explain the main concepts and vocabulary introduced throughout the volume.
- Apply a structured method to a relevant practical problem.
- Interpret evidence, identify limitations, and communicate results clearly.
- Consider privacy, fairness, safety, security, and human oversight where relevant.
How to use this book
Read the chapter explanation, complete the activity, answer the review questions, and keep a learning journal recording assumptions, decisions, mistakes, and improvements. For technical work, save code, data dictionaries, test results, and software-version details.
Table of contents
- Chapter 1: Mathematical Foundations of Deep Learning
- Chapter 2: Artificial Neurons and Neural Networks
- Chapter 3: Loss Functions and Optimisation
- Chapter 4: Backpropagation and Gradient-Based Learning
- Chapter 5: Convolutional Neural Networks
- Chapter 6: Sequence Models and Attention
- Chapter 7: Transformers and Large Language Models
- Chapter 8: Generative AI and Foundation Models
- Chapter 9: Prompting, Retrieval, and Grounded Responses
- Chapter 10: Embeddings and Semantic Search
- Chapter 11: Fine-Tuning and Adaptation
- Chapter 12: Model Evaluation and Hallucination Analysis
- Chapter 13: Multimodal AI Systems
- Chapter 14: Deployment, Inference, and Performance
- Chapter 15: Security, Privacy, and Responsible Generative AI
- Chapter 16: Advanced Application Project
Chapter 1: Mathematical Foundations of Deep Learning
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in mathematical foundations of deep learning.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Deep learning relies on linear algebra, calculus, probability, and optimisation. Vectors and matrices represent data and parameters; derivatives describe how a loss changes as parameters change.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Work through a small vector calculation and explain what a parameter represents.
Chapter review
- What is the main idea in 'Mathematical Foundations of Deep Learning'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 2: Artificial Neurons and Neural Networks
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in artificial neurons and neural networks.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Artificial neurons combine weighted inputs, a bias, and an activation function. Layers build representations, but architecture should be justified by task and data rather than complexity alone.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Draw a small feed-forward network and trace one input through it conceptually.
Chapter review
- What is the main idea in 'Artificial Neurons and Neural Networks'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 3: Loss Functions and Optimisation
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in loss functions and optimisation.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
A loss function quantifies model error according to a chosen objective. Optimisers update parameters; learning rate, batch size, regularisation, and data quality affect training.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Compare two hypothetical training curves and diagnose underfitting, overfitting, or unstable learning.
Chapter review
- What is the main idea in 'Loss Functions and Optimisation'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 4: Backpropagation and Gradient-Based Learning
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in backpropagation and gradient-based learning.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Backpropagation applies the chain rule to compute gradients through a network. Gradient-based optimisers use those gradients to reduce loss, but optimisation success does not guarantee useful generalisation.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Explain the direction of a gradient update in a one-parameter example.
Chapter review
- What is the main idea in 'Backpropagation and Gradient-Based Learning'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 5: Convolutional Neural Networks
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in convolutional neural networks.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Convolutional neural networks use local filters and shared parameters to learn spatial features in images. Pooling, augmentation, and architecture choices affect accuracy, computation, and robustness.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Design an image-classification experiment and identify likely sources of dataset leakage.
Chapter review
- What is the main idea in 'Convolutional Neural Networks'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 6: Sequence Models and Attention
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in sequence models and attention.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Sequence models process ordered data such as text or time series. Attention weights relationships among elements and supports flexible contextual representations.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Explain with a short sentence how attention might relate one word to another.
Chapter review
- What is the main idea in 'Sequence Models and Attention'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 7: Transformers and Large Language Models
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in transformers and large language models.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Transformers use attention and related components to process token sequences. Language models are trained on data to predict or otherwise model token sequences; generated fluency does not guarantee factual accuracy.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Trace the high-level pipeline from prompt tokens to generated text and identify uncertainty points.
Chapter review
- What is the main idea in 'Transformers and Large Language Models'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 8: Generative AI and Foundation Models
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in generative ai and foundation models.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Foundation models are trained on broad datasets and adapted to many downstream tasks. Generative systems produce new outputs conditioned on inputs; quality, provenance, rights, and safety need consideration.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Compare a generative model use case with a conventional classifier use case.
Chapter review
- What is the main idea in 'Generative AI and Foundation Models'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 9: Prompting, Retrieval, and Grounded Responses
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in prompting, retrieval, and grounded responses.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Prompts specify task and constraints; retrieval systems supply relevant source material; grounded generation attempts to tie responses to that material. Retrieval can still miss relevant sources or surface misleading content.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Create a prompt and source-grounded answer rubric that requires evidence and an uncertainty statement.
Chapter review
- What is the main idea in 'Prompting, Retrieval, and Grounded Responses'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 10: Embeddings and Semantic Search
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in embeddings and semantic search.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Embeddings map items into vector representations where distance can approximate selected semantic relationships. Similarity depends on model, corpus, chunking, and task; nearest neighbours are not guaranteed to be correct.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Design a small semantic-search evaluation set with relevant and irrelevant examples.
Chapter review
- What is the main idea in 'Embeddings and Semantic Search'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 11: Fine-Tuning and Adaptation
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in fine-tuning and adaptation.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Adaptation may include prompting, retrieval augmentation, parameter-efficient tuning, or full fine-tuning. Choose based on task, data rights, budget, latency, and maintenance needs.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Prepare a decision table comparing retrieval augmentation and fine-tuning for a document assistant.
Chapter review
- What is the main idea in 'Fine-Tuning and Adaptation'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 12: Model Evaluation and Hallucination Analysis
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in model evaluation and hallucination analysis.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Evaluation needs representative tasks, explicit criteria, human review, and tests for failure modes. Hallucination analysis examines unsupported claims, not just grammatical quality or user satisfaction.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Create a test set with factual, ambiguous, out-of-scope, and adversarial questions.
Chapter review
- What is the main idea in 'Model Evaluation and Hallucination Analysis'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 13: Multimodal AI Systems
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in multimodal ai systems.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Multimodal systems combine or translate between text, image, audio, video, or structured signals. Modality-specific privacy, bias, and error patterns need separate evaluation.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Map the data flows and privacy risks of a fictional image-and-text assistant.
Chapter review
- What is the main idea in 'Multimodal AI Systems'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 14: Deployment, Inference, and Performance
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in deployment, inference, and performance.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Deployment requires model serving, version control, latency and throughput monitoring, security, rollback, and cost management. Quantisation or caching can change performance and output quality.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Draft a deployment checklist covering observability, rollback, access, and evaluation.
Chapter review
- What is the main idea in 'Deployment, Inference, and Performance'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 15: Security, Privacy, and Responsible Generative AI
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in security, privacy, and responsible generative ai.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
Risks include prompt injection, data leakage, unsafe outputs, misuse, and overreliance. Controls include input/output filtering, least privilege, isolation, testing, and human review.
Worked example and application
Worked example: imagine a small team considering this concept for a clearly scoped, low-risk task. First define the intended outcome and inputs; then choose a simple baseline, document assumptions, test representative examples, inspect errors, and explain when the method should not be used. Compare the result with a human-reviewed alternative before proposing deployment.
Practical activity
Threat-model a retrieval-augmented assistant and identify controls for untrusted documents.
Chapter review
- What is the main idea in 'Security, Privacy, and Responsible Generative AI'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Chapter 16: Advanced Application Project
Learning outcomes
By the end of this chapter, the learner should be able to:
- Explain the central concepts introduced in advanced application project.
- Apply the ideas to a clearly defined example or practical task.
- Identify at least one limitation, assumption, or responsible-use consideration.
Core explanation
A practical system needs a defined use case, source or training data, architecture, evaluation, risk controls, and a plan for maintenance. A demo alone does not establish production readiness.
Worked example and application
Worked example: define a real-world task, list the inputs and constraints, sketch the steps or system components, and test the proposed solution with at least one ordinary case and one edge case. Record the outcome and what you would improve.
Practical activity
Build and evaluate an AI-assisted knowledge application using a defined source collection, with clear limits and human review.
Chapter review
- What is the main idea in 'Advanced Application Project'?
- Which assumptions or data requirements matter most in this topic?
- What could cause the approach to fail or produce a misleading result?
- How would you explain the result or limitation to a non-specialist?
Key takeaways
- Start with a clear question, objective, or system requirement.
- Document inputs, assumptions, methods, and limitations.
- Test the approach with relevant examples and review the consequences of errors.
Capstone project
Build and evaluate an AI-assisted knowledge application using a defined source collection. Document architecture, data permissions, retrieval or adaptation choices, evaluation set, unsupported-answer behaviour, security controls, and human review.
Suggested project deliverables
- A short problem statement and intended audience.
- A description of data, inputs, assumptions, tools, and method.
- Evidence of testing, analysis, or evaluation appropriate to the project.
- A clear explanation of results, limitations, and foreseeable risks.
- A final report or presentation and a reflection on possible improvements.
Assessment guidance
Evaluate technical reasoning, appropriate model selection, robust evaluation, security threat modelling, reproducibility, and honest reporting of limitations.
Glossary
Embedding: A numerical representation intended to preserve selected relationships between items.
Transformer: A neural-network architecture built around attention mechanisms.
Fine-tuning: Further training a pretrained model on selected data or tasks.
Retrieval-augmented generation: A system pattern that retrieves external information to support generated responses.
Hallucination: A generated claim that is unsupported or incorrect relative to the required evidence.
Responsible practice checklist
- Use data and materials only when their source and permitted use are understood.
- Avoid exposing personal, confidential, or credential information.
- Document assumptions, missing information, and important limitations.
- Test ordinary cases, boundary cases, and foreseeable failure conditions.
- Use human review where errors could materially affect people, safety, rights, or resources.
- Do not claim that a system is fair, safe, secure, compliant, or reliable without appropriate evidence.
References and further reading
The source material supplied for this project does not provide a complete, volume-specific bibliography. Before publication, add verified primary sources, textbooks, official documentation, and standards applicable to this volume. Confirm exact titles, editions, publication dates, URLs, licences, and access dates. Do not cite a standard as evidence of conformity unless the required assessment has actually been performed.
[1] C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2006.
[2] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, MA, USA: MIT Press, 2016.
[3] National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, Jan. 2023.
[4] IEEE, “IEEE Author Center.” [Online]. Available: https://journals.ieeeauthorcenter.ieee.org/
[5] IEEE Standards Association, “IEEE Standards.” [Online]. Available: https://standards.ieee.org/