EXIN Artificial Intelligence Foundation

• Human Intelligence -- "The mental quality that consists of the abilities to learn from experience, adapt to new situations, understand and handle abstract concepts, and use knowledge to manipulate one's environment."
• Artificial Intelligence (AI) -- "Intelligence demonstrated by machines, in contrast to the natural intelligence displayed by humans and other animals."
• Machine Learning (ML) -- "The study of computer algorithms that allow computer programs to automatically improve through experience."
• Scientific Method -- "An empirical method for acquiring knowledge that has characterized the development of science."

Key elements of the scientific method: Observation, Hypothesis, Experimentation, Analysis, Replication, Peer Review.

The exam tests five key milestones only:

• Asilomar Principles (2017) -- Coordinated by the Future of Life Institute (FLI) at the Beneficial AI 2017 conference. A set of 23 principles covering three categories: Research issues, Ethics and values, Longer-term issues.
• Dartmouth Conference (1956) -- Birthplace of AI. The term "Artificial Intelligence" was coined here. Organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon.
• AI Winters -- Periods of reduced funding and interest:
  • First AI Winter (1974-1980): Overpromises, underperformance, funding cuts.
  • Second AI Winter (1987-1993): Commercial failures of expert systems, overhype, computational limitations.
• Big Data and IoT -- Enormous datasets from social media, sensors, e-commerce, and IoT devices fueled modern AI advances.
• Large Language Models (LLMs) -- Widespread public use from 2022 onward.

• Narrow AI (ANI) -- Also known as weak AI. Task-specific, operates within well-defined domains. All current AI systems are narrow AI. Examples: image recognition, speech recognition, language translation, virtual assistants (Siri, Alexa), spam filtering, medical diagnostics, generative AI.
• General AI (AGI) -- Also known as strong AI. Aims to replicate human intelligence. The hypothetical intelligence of a machine that has the capacity to understand or learn any intellectual task a human can. AGI does not currently exist.

Floridi & Cowls' Principles (5 principles for ethical AI):

1. Beneficence -- AI should promote well-being and do good
2. Non-maleficence -- AI should avoid causing harm ("do no harm")
3. Autonomy -- AI should respect human autonomy and decision-making
4. Justice -- AI should promote fairness and avoid reinforcing inequality
5. Explicability -- AI systems should be transparent and understandable

UK AI Principles (5 principles):

1. Safety, security and robustness
2. Transparency and explainability
3. Fairness
4. Accountability and governance
5. Contestability and redress

• Social impact: Job displacement vs. creation, reskilling needs, privacy concerns
• Economic impact: Productivity gains, innovation, economic disparities
• Environmental impact: Energy consumption, water usage, carbon emissions, e-waste
• UN 17 Sustainable Development Goals (SDGs) -- AI can help or hinder these
• EU AI Act (2024) -- Regulatory framework for AI in the EU

• Green IT initiatives -- Energy-efficient hardware, renewable energy, responsible disposal
• Data center energy and efficiency -- Renewable energy, improved cooling, optimized utilization
• Sustainable supply chain -- Responsible sourcing, reduced manufacturing emissions, minimized e-waste
• Choice of algorithm -- More efficient algorithms reduce energy use; not every problem needs deep learning
• Low-code/no-code programming -- Reduces development overhead and resource consumption
• Monitoring and reporting environmental impact -- Tracking energy, carbon, and water use throughout the AI lifecycle

• Ethics -- "Moral principles that govern a person's behaviour or the conducting of an activity." (Oxford English Dictionary)
• Ethics vs. Law: Ethics = moral guidelines (can vary); Law = formal rules enforced by government. Overlap exists, but they are distinct.

Ethical concerns in AI:

• Bias, unfairness, and discrimination -- AI can perpetuate or amplify biases from training data
• Data privacy and protection -- AI processes vast personal data, raising privacy concerns
• Impact on employment and the economy -- Automation replacing jobs, widening inequality
• Autonomous weapons -- Ethical issues with lethal autonomous systems
• Autonomous vehicles and liability -- Who is responsible when self-driving cars cause harm?

The UK AI Principles guide ethical AI:

1. Safety, security and robustness
2. Transparency and explainability
3. Fairness
4. Accountability and governance
5. Contestability and redress

AI Governance -- A set of practices to keep AI systems under control so they remain safe and ethical. Includes organizational policies, standards, and AI steering committees.

Ethical challenges (threats to ethical behavior):

• Self-interest -- Placing personal gain above ethical obligations
• Self-review -- Reviewing your own work without independent scrutiny
• Conflict of interest -- Competing loyalties that compromise objectivity
• Intimidation -- Being pressured to act unethically
• Advocacy -- Promoting a position to the point of compromising objectivity

Strategies for addressing challenges:

• Dealing with bias -- Diverse data, diverse teams, fairness metrics
• Openness -- Transparency about AI use, data sources, and limitations
• Transparency -- Making AI processes visible and understandable
• Trustworthiness -- Building reliable, dependable systems
• Explainability -- AI decisions must be explainable in human terms

An ethical risk framework integrates ethical considerations into every stage of AI development.

• Need for regulation -- Ensures legal accountability and effective management of AI
• AI regulation landscape -- Standards like WCAG (Web Content Accessibility Guidelines)
• Data Protection Act 2018 (DPA 2018) -- UK data protection legislation
• UK GDPR -- Governs personal data collection, storage, and processing in the UK
• ISO -- International standards for AI systems
• NIST -- US frameworks for AI risk management
• Consequences of unregulated AI -- Widespread harm, bias, loss of trust, privacy violations
• Professional standards -- Must be ethical, accountable, competent, inclusive

• Risk -- "A person or thing regarded as a threat or likely source of danger."
• Risk management -- "A process or series of processes which allow risk to be understood and minimized proactively."

Risk management techniques:

• Risk analysis -- Identifying and assessing potential risks
• SWOT analysis -- Strengths, Weaknesses, Opportunities, Threats
• PESTLE analysis -- Political, Economic, Social, Technological, Legal, Environmental
• Cynefin framework -- Categorizes problems: Simple/Clear, Complicated, Complex, Chaotic, Disorder

Risk mitigation strategies:

• Ownership and accountability -- Assigning clear risk owners
• Stakeholder involvement -- Engaging all affected parties
• Subject matter experts -- Consulting domain experts

• Human compatible -- AI working alongside humans (cobots, collaborative tools)
• Wearable -- Fitness trackers, smartwatches with health monitoring
• Edge -- AI processing at the device level (faster, more private)
• Internet of Things (IoT) -- Smart home devices, sensors, connected equipment
• Personal care -- AI health apps, personalized medication, mental health chatbots
• Self-driving vehicles -- Autonomous cars using sensors, cameras, and AI
• Generative AI tools -- ChatGPT, image generators, code assistants

• Robotics -- "A machine that can carry out a complex series of tasks automatically, either with or without intelligence."
• Intelligent vs. non-intelligent: Intelligent robots use AI (sensors, learning); non-intelligent follow fixed instructions.

Types of robots:

• Industrial -- Manufacturing, assembly, welding
• Personal -- Home assistants, vacuum cleaners
• Autonomous -- Self-driving vehicles, drones
• Nanobots -- Microscopic robots for medical applications
• Humanoids -- Robots resembling humans (ASIMO, Sophia)

Robotic Process Automation (RPA) -- Software that automates repetitive digital tasks (data entry, form processing). Not a physical robot.

• Machine Learning -- "The field of machine learning is concerned with the question of how to construct computer programs that automatically improve with experience." (Tom Mitchell)
• Neural Networks -- "A machine learning program, or model, that makes decisions in a manner similar to the human brain, by using processes that mimic the way biological neurons work together to identify phenomena, weigh options and arrive at conclusions."
• Deep Learning -- "Deep learning is a multi-layered neural network."
• Large Language Models (LLMs) -- "LLMs are deep learning algorithms that can recognize, summarize, translate, predict, and generate content using very large datasets." (IBM)

Hierarchy: AI > Machine Learning > Deep Learning > LLMs

• Prediction -- Using historical data to forecast future outcomes
• Object recognition -- Identifying objects within images or video (CNNs)
• Classification -- Assigning data to predefined categories; includes random decision forests (ensemble of decision trees)
• Clustering -- Grouping data by similarities without predefined categories (unsupervised)
• Recommendations -- Suggesting content based on behavior (e.g., Netflix, Spotify)

• Supervised learning -- Uses labeled data (input-output pairs). We know what the output will be. Example: spam classification.
• Unsupervised learning -- Uses unlabeled data. Discovers hidden patterns. Example: customer segmentation/clustering.
• Semi-supervised learning -- Small amount of labeled data + larger amount of unlabeled data. Useful when labeling is expensive.

• Big Data -- "Extremely large data sets that may be analyzed computationally to reveal patterns, trends, and associations." (Dialogic.com)
• Data Visualization -- "The representation of data through use of common graphics, such as charts, plots, infographics and even animations." (IBM)
• Structured data -- Data organized sequentially or serially in a tabular format. Examples: spreadsheets, SQL databases.
• Semi-structured data -- Has some organizational properties but not tabular. Examples: JSON, XML, emails.
• Unstructured data -- No pre-defined order or structure. Examples: images, videos, social media posts.

Five data quality characteristics:

1. Accuracy -- Is the data correct?
2. Completeness -- Is all the required data present?
3. Uniqueness -- Is the data free from duplication?
4. Consistency -- Is the data free from conflict across sources?
5. Timeliness -- Is the data current and available when needed?

Implications of poor-quality data:

• Errors and inaccuracies in AI outputs
• Bias amplified through flawed data
• Loss of trust in AI systems
• Financial penalties from non-compliance

• Bias: Multiple sources, diversity in teams handling data, fairness metrics
• Misinformation: Check reliability of sources, use subject matter expert (SME) validation
• Processing restrictions: Organizational requirements, frameworks and regulations
• Legal restrictions: UK GDPR, DPA 2018, staying abreast of new requirements
• The scientific method: Hypothesis-driven, evidence-based approach for AI development

• Storage and use -- Cost-effective storage and processing of massive datasets
• Understanding the user -- Analyzing behavior and preferences for targeted marketing
• Improving process -- Identifying inefficiencies, data-driven business decisions
• Improving experience -- Personalization and predictive analytics for better user experiences

• Written -- Reports, summaries, narrative descriptions
• Verbal -- Presentations, spoken explanations
• Pictorial -- Charts, graphs, plots, diagrams, maps
• Sounds -- Audio alerts, sonification
• Dashboards and infographics -- Interactive multi-visualization displays
• Virtual and augmented reality (VR/AR) -- Immersive 3D environments

• Generative AI -- "Refers to deep-learning models that can generate high-quality text, images, and other content based on the data they were trained on." (IBM)
• LLMs -- "Deep learning algorithms that can recognize, summarize, translate, predict, and generate content using very large datasets." (IBM)

• Trained on huge volumes of data
• Uses training to predict the next word in a chain of words
• Generates coherent, human-sounding language
• Prompt engineering -- Crafting specific, detailed requests for better AI outputs
• Natural Language Processing (NLP) -- Machines understanding and processing human language
• Image generation -- Creating images from text descriptions

Stages of the ML training process (in order):

1. Analyze the problem -- Define what you are solving
2. Data selection -- Choose relevant data sources
3. Data pre-processing -- Clean, normalize, transform data
4. Data visualization -- Explore and understand patterns
5. Select a model/algorithm -- Choose the appropriate ML approach
   • Train the model
   • Test the model
   • Repeat (learn from experience)
6. Review -- Assess the overall solution

• Automation -- Automating processes to minimize human input
• Repetitive tasks -- AI handles routine work humans find tedious
• Content creation via generative AI -- Drafting text, generating images, producing reports

1. Introduction -- Context and purpose
2. Management or executive summary -- High-level overview
3. Description of current state -- How things work today
4. Options considered -- Each includes:
   • Option described
   • Analysis of costs and benefits
   • Impact assessment
   • Risk assessment
5. Recommendations -- Proposed course of action
6. Appendices/supporting information -- Detailed data, references

Stakeholder -- Any individual or group with an interest in or influence on a project.

Power/Interest Grid (4 quadrants):

• High power, High interest -- Constant active management (key players)
• High power, Low interest -- Keep satisfied
• Low power, High interest -- Keep informed
• Low power, Low interest -- Monitor

Stakeholder Wheel -- Visual representation of all stakeholder groups around a project.

• Agile -- Iterative, flexible, sprints, embraces changing requirements, continuous feedback
• Waterfall -- Sequential, linear. Phases: Requirements > Design > Implementation > Testing > Deployment > Maintenance
• Hybrid -- Combines Agile and Waterfall elements

• Risk analysis: Risk assessment + risk owners
• Risk appetite -- Level of risk the organization will accept
• Risk management strategies:
  • Accept -- Acknowledge, take no action
  • Mitigate -- Reduce probability/impact (inc. sharing and contingency planning)
  • Avoid -- Change plans to eliminate the risk
  • Transfer -- Shift risk to another party (e.g., insurance)
• Financial costs/benefits: Forecasting, margin for error
• Socio-economic benefits -- Wider societal gains
• Triple bottom line:
  • Profit -- Financial performance
  • People -- Social impact
  • Planet -- Environmental sustainability

Three governance areas:

• Compliance -- Satisfying all applicable regulations
• Risk management -- Proactively detecting and mitigating risks
• Lifecycle governance -- Ongoing management:
  • Manage -- Day-to-day operations
  • Monitor -- Continuous tracking of performance, bias, drift
  • Govern -- Strategic oversight and policy enforcement

AI-specific roles:

• Machine Learning Engineer -- Designs and builds ML models
• Data Scientist -- Analyzes complex data, builds predictive models
• AI Research Scientist -- Conducts fundamental AI research
• Computer Vision Engineer -- Interprets visual data
• NLP Engineer -- Builds language understanding systems
• Robotics Engineer -- Designs robots and autonomous systems
• AI Ethics Specialist -- Ensures responsible AI development
• AI Anthropologist -- Studies cultural/social implications of AI

Opportunities for existing roles: additional training and knowledge, improved efficiency, automation of routine tasks.

• Marketing -- Trend prediction, targeted advertising, customer segmentation
• Healthcare -- Diagnostics (X-rays, MRIs), treatment planning, drug discovery
• Finance -- Fraud detection, algorithmic trading, credit scoring, audit automation
• Transportation -- Self-driving cars, route optimization, traffic management
• Education -- Personalized learning, adaptive assessments
• Manufacturing -- Predictive maintenance, quality control, supply chain optimization
• Entertainment -- Recommendation algorithms (Netflix, Spotify), content generation
• IT -- Cybersecurity, chatbots, automated testing, infrastructure management

Benefits:

• Reducing human error through automation
• Processing vast data for informed decisions
• AI-powered medical diagnosis assistance

Challenges:

• Algorithm bias and privacy concerns
• Job loss and displacement
• Security risks from hacking
• Socio-economic inequality
• Lack of creativity and empathy

Environmental impact: Energy consumption, climate change, e-waste.

Economic impact: Job losses in some sectors, need for retraining, market volatility.

Future advancements: Increased computing power, more data, better algorithms.

• Human consciousness (sentience) -- Subjective experience of awareness; capacity for feelings, perceptions, self-awareness
• AI consciousness -- Hypothetical: could AI develop genuine subjective experience?
• Kurzweil Singularity -- "A future period characterized by rapid technological growth that will irreversibly transform human life."
• Seth's theory (Anil Seth):
  • Predictive processing and perception -- The brain constructs reality through predictions
  • The nature of self and consciousness -- Consciousness is about prediction, not just information processing
• Functional capabilities vs. genuine consciousness -- AI may mimic conscious behavior without actually being conscious
• Ethical implications -- Should AI appear human? If AI were conscious, would it have rights?