Applying Artificial Intelligence: The Practitioner's Handbook

✍ Scribed by Rose Johansen, Dan

• Publisher: Biased Publications
• Year: 2024
• Tongue: English
• Leaves: 331
• Category: Library

✦ Synopsis

On the back of the cover
Although AI technology has become mature, affordable, and robust and promises immense business value, implementations often fail with broken budgets and dispirited staff in its wake.

According to the author, the problem is always the same. The project methods we use for designing and implementing traditional IT do not work for AI. Thus, the high failure rate is not caused by the technology but by how we run the projects.

In the book, the author explains how AI differs from traditional IT. He then provides an easy-to-follow step-by-step method to identify the main challenges, write the business case, plan the project, design the solution, get the users on board, and deliver tangible business value.

✦ Table of Contents

Applying Artificial Intelligence
Title Page
Copyright
Dedication
Contents
Foreword by Lars Tvede
Introduction
1. Making Book-keeping More Productive
Chaotic first year
First and confusing feedback
Beta-launch
Humans in the loops – what we did right
Growth
Research and development don’t always pay off
Going global
Happy ending
2. Defining and Demystifying AI
Machine learning
Artificial intelligence – a misleading name
Defining AI is essential
Cat or dog?
Data, information and predictions
Relations and casualties
Three (or so) ways for AI to learn from data
Generative versus predictive AI
Generative AI can also predict
3. AI Building Blocks
No homemade pasta
Building blocks
Labels and features are fundamental to AI building blocks.
Tabular building blocks
Classification
Regression
Forecasting
Vision building blocks
Classification
Object detection
Image segmentation
Image similarity
Language Building blocks
Classification
Named entity recognition
Intent analysis
Sentiment analysis
Summarising text
Sound
Audio classification
Multimodal building blocks
4. Preparing Your Organisation for Reaping the Benefits of AI
Domain experts
Knowledge drives demand
It's not about hiring technicians first
Why the encapsulated and technical approach fails
AI is bottom-up
How do you teach AI?
Bridge the gap of ignorance.
To adopt or not to adopt isn't a choice
A technology for solving business problems
Where is your organisation today?
Ignorant
Three adoption strategies
Summary
5. Solving Problems Using AI
Methods aren’t new
The Todai Method
Method steps
Using the method
1. Inspiration phase
2. Discovery
3. Data handling
4. Development
5. Implementation
6. Monitoring
Milestone funding
Mind the mindset
Building AI is like developing a product
How to be product-focused
Measure user adoption and make it your responsibility
Think like a product manager
Build a product vision
Find product market fit
User experience design
Summary
6. Inspiration Phase
Objectives
Activities
Training people to identify opportunities
Make AI accessible
Idea workshop
Add a pretotype
Design sprints
Pre-analysis
Decision models
Ask for data
Build models
Interview domain experts
Look what others have done
Inspiration phase checklist
Building the business case
Moving forward
Set goals and calculate backwards
Assume everything will go wrong
Secure stakeholder support
Summary
7. Discovery – Understanding the Problem
Discovery phase
Discovery phase activities
Solving the right problem
What are you trying to achieve?
Selecting a business outcome
The Paperflow problem
The SundAI problem
Connecting with strategy
Understand the problem domain
Problem components
Understanding the Paperflow problem
Use well-known frameworks
Has the problem been solved before?
Do you agree on the problem?
Use a workshop tool
Acquiring domain knowledge through subject matter experts
Data is only half the story
Domain expert interviews
Exploring feature engineering
Implicit knowledge
Beware of bias
Interview more experts
Expertise-induced blindness
Interview guide
Testing before building
Summary
8. Discovery – Decision Making and Accuracy
The effort in AI is skewed
Decision model
Generative models and decision models
Confidence thresholds
Decision models incite new models
You probably want more info
Other types of models
Decision levels
Level 1: Warn/inform
Level 2: Suggest
Level 3: Conditional automation
Level 4: Automatic
Level 5: Autonomous
Confidence and accuracy
Accuracy and confidence are not the same
Accuracy value
Agree on how you measure accuracy
Not all information is equal in value
Perceived accuracy
Who are you trying to beat?
The value is in the decision
Decisions first – always
Build the business case on decisions
Decision flows
Element of luck
Map the flow
Decision-making is complicated
Decisions are feelings
Kahneman has a point
Decisions are affected by politics, compliance and strategy
Summary
9. Data handling
Activities in the data handling phase
Goal of the data handling phase
Data terminology
The goal of data is to represent the world in which the AI should work
Perfection is impossible
How to handle it
Data quality
User-generated data
How much data and at what cost?
Building blocks matter the most for collection costs
Generative AI data costs
Adoption strategies and data costs
Foundation models
The desired accuracy
Optimal cost
Faster and better data labelling
Model-assisted labelling
Active learning
Understanding and analysing data
Look at the data
Black swans
Data bias
Mitigating bias
Data as a competitive advantage
Personal data
Challenges
Handling PII
Synthetic data
Add data to labels with no or little data
Remove personal data
Whether data is synthetic depends on the use
Synthetic data will become a big deal
Types of synthetic data
Synthetic texts
Synthetic images
Synthetic tabular data
Models of the world
Summary
10. Developing AI
The development phase activities
Team roles in AI projects and their challenges
Still a need for data scientists?
Anatomy of AI solutions
Model(s) and prediction pipeline
Generative models
Mindset for building AI
Experiment efficiently
End-to-end first
Be ready to start over within twenty-four hours
Speed is of the essence
Simplest approach possible
Building and understanding models
Understanding models
Explainable AI
Testing results
Monitoring phase: Building AI operations
Monitoring AI activities
Operations, feedback, and retraining
11. Implementation phase – AI and Humans
What makes implementation so important?
Starting the implementation phase
Before the implementation phase
The implementation phase reveals discovery failures
Achieving adoption
You get what you measure
Adoption is a marketing task
Humans find patterns
Gaining trust or treating fears and insecurities
The problem
Solution
Sense of control
The Kasparov perspective on AI adoption
Social contract
Summary
Afterword
Acknowledgments
About the Author
Index
Notes