AI Terminology and Concepts

To effectively implement an AI Management System, you need to understand fundamental AI concepts and terminology. This lesson provides the foundation of AI knowledge necessary for ISO 42001 implementation.

What is Artificial Intelligence?

Defining AI

ISO/IEC 22989 Definition: An engineered system that generates outputs such as content, forecasts, recommendations, or decisions for a given set of human-defined objectives.

Practical Definition: AI systems perform tasks that typically require human intelligence, such as:

• Recognizing patterns
• Making decisions
• Understanding language
• Perceiving environments
• Learning from experience
• Solving complex problems

Characteristics of AI Systems

Autonomy: AI systems operate with some degree of independence, making decisions without constant human direction.

Adaptability: AI systems can adjust behavior based on data, feedback, or changing conditions.

Goal-Oriented: AI pursues objectives, whether explicitly programmed or learned.

Data-Driven: AI learns from data rather than following only predefined rules.

Probabilistic: AI typically provides probabilistic outputs (confidence levels, likelihoods) rather than absolute certainty.

AI vs. Traditional Software

Core AI Concepts

Machine Learning (ML)

Definition: A subset of AI where systems learn patterns from data without being explicitly programmed for every scenario.

How It Works:

1. System receives training data (examples)
2. Algorithm finds patterns in the data
3. System creates a model representing those patterns
4. Model makes predictions or decisions on new data
5. Performance is evaluated and model may be refined

Types of Machine Learning:

Supervised Learning: Algorithm learns from labeled examples (input-output pairs).

Example: Training spam filter with emails labeled "spam" or "not spam"

Common Algorithms: Decision trees, random forests, support vector machines, neural networks

Applications: Classification, regression, prediction

Unsupervised Learning: Algorithm finds patterns in unlabeled data without predefined categories.

Example: Customer segmentation discovering groups of similar customers

Common Algorithms: K-means clustering, hierarchical clustering, principal component analysis (PCA)

Applications: Clustering, anomaly detection, dimensionality reduction

Reinforcement Learning: Agent learns by interacting with environment, receiving rewards or penalties.

Example: Game-playing AI learning optimal strategies through trial and error

Common Algorithms: Q-learning, deep Q-networks (DQN), policy gradient methods

Applications: Robotics, game playing, autonomous systems, optimization

Semi-Supervised Learning: Combines small amount of labeled data with large amount of unlabeled data.

Use Case: When labeling is expensive or time-consuming

Applications: Medical imaging (limited expert annotations), text classification

Deep Learning (DL)

Definition: Subset of machine learning using artificial neural networks with multiple layers (deep neural networks).

Key Characteristics:

• Multiple processing layers extracting increasingly abstract features
• Automatic feature engineering (learns relevant features from raw data)
• Requires large amounts of data and computational power
• Particularly effective for complex patterns in images, text, and audio

Neural Network Basics:

• Neurons: Basic processing units that receive inputs, apply weights, and produce outputs
• Layers: Input layer (receives data), hidden layers (process information), output layer (produces results)
• Weights: Parameters learned during training, determining how inputs influence outputs
• Activation Functions: Introduce non-linearity, enabling networks to learn complex patterns

Common Deep Learning Architectures:

Convolutional Neural Networks (CNNs):

• Specialized for processing grid-like data (images)
• Use convolutional layers to detect local patterns
• Applications: Image recognition, computer vision, medical imaging

Recurrent Neural Networks (RNNs):

• Process sequential data by maintaining internal state
• Handle variable-length sequences
• Applications: Language modeling, time series prediction, speech recognition

Transformers:

• Process sequences using attention mechanisms
• Highly parallelizable and effective for long-range dependencies
• Foundation for modern language models
• Applications: Natural language processing, machine translation, large language models

Generative Adversarial Networks (GANs):

• Two networks (generator and discriminator) competing
• Generator creates synthetic data; discriminator distinguishes real from fake
• Applications: Image generation, data augmentation, style transfer

Natural Language Processing (NLP)

Definition: AI techniques enabling computers to understand, interpret, and generate human language.

Key Capabilities:

• Text classification (sentiment analysis, topic categorization)
• Named entity recognition (identifying people, places, organizations)
• Machine translation (translating between languages)
• Question answering (providing answers to natural language questions)
• Text generation (creating coherent text)
• Summarization (condensing text while preserving key information)

Modern NLP Approaches:

Word Embeddings: Representing words as dense vectors capturing semantic relationships.

Example: Word2Vec, GloVe enabling "king - man + woman ≈ queen"

Language Models: Models predicting next word or token in sequence.

Traditional: N-gram models based on word frequencies

Modern: Transformer-based models like GPT, BERT

Large Language Models (LLMs): Massive transformer models trained on vast text corpora.

Examples: GPT-4, Claude, Gemini, LLaMA

Capabilities: Complex reasoning, multi-task performance, in-context learning

Computer Vision

Definition: AI enabling computers to understand and interpret visual information from the world.

Key Tasks:

Image Classification: Categorizing entire image into predefined classes.

Example: Identifying whether image contains a cat or dog

Object Detection: Locating and classifying multiple objects within image.

Example: Autonomous vehicles detecting pedestrians, traffic signs, other vehicles

Semantic Segmentation: Classifying each pixel in image.

Example: Medical imaging identifying tumor boundaries

Image Generation: Creating new images based on descriptions or examples.

Example: Text-to-image systems like DALL-E, Midjourney, Stable Diffusion

Facial Recognition: Identifying or verifying individuals from facial features.

Example: Phone unlocking, security systems

Applications: Medical diagnosis, autonomous vehicles, security surveillance, quality control, augmented reality

Generative AI

Definition: AI systems that create new content (text, images, audio, video, code) rather than just analyzing or classifying existing data.

How It Works:

• Trained on large datasets of examples
• Learns patterns and relationships in data
• Generates new content that resembles training data but is novel

Types of Generative AI:

Large Language Models: Generate coherent text for various purposes.

Examples: ChatGPT, Claude, Gemini Uses: Writing assistance, code generation, conversation, analysis

Text-to-Image: Generate images from text descriptions.

Examples: DALL-E, Midjourney, Stable Diffusion Uses: Art creation, design, prototyping, visualization

Text-to-Video: Generate video from text prompts.

Examples: Sora, Runway, Synthesia Uses: Content creation, education, marketing

Audio Generation: Generate music, speech, or sound effects.

Examples: MusicLM, Jukebox, ElevenLabs Uses: Music composition, voice synthesis, sound design

Code Generation: Generate software code from natural language descriptions.

Examples: GitHub Copilot, CodeLlama, Amazon CodeWhisperer Uses: Programming assistance, automation, education

Significance for ISO 42001: Generative AI raises unique governance challenges around content authenticity, intellectual property, misinformation, and responsible use.

AI System Components

Data

Training Data: Data used to train AI model, teaching it patterns and relationships.

Validation Data: Data used during training to tune model parameters and prevent overfitting.

Test Data: Data held out until final evaluation, measuring model's performance on unseen examples.

Inference Data: New data the trained model processes to make predictions or decisions.

Data Quality Factors:

• Accuracy: Data correctly represents reality
• Completeness: All necessary data is present
• Consistency: Data is uniform and not contradictory
• Timeliness: Data is current and up-to-date
• Representativeness: Data reflects the population model will encounter
• Relevance: Data is pertinent to the task

Models

Definition: Mathematical representation of patterns learned from training data.

Types:

• Predictive Models: Forecast future values or outcomes
• Classification Models: Assign inputs to predefined categories
• Generative Models: Create new content or data
• Descriptive Models: Identify patterns and relationships
• Prescriptive Models: Recommend actions to achieve objectives

Model Parameters: Values learned during training (e.g., neural network weights).

Hyperparameters: Configuration choices made before training (e.g., learning rate, number of layers).

Model Architecture: Structure and design of the model (e.g., network topology).

Algorithms

Training Algorithms: Methods for learning model parameters from data.

Examples: Gradient descent, backpropagation, stochastic optimization

Inference Algorithms: Methods for using trained model to process new data.

Examples: Forward propagation in neural networks, tree traversal in decision trees

Optimization Algorithms: Techniques for improving model performance.

Examples: Adam, RMSprop, L-BFGS

Infrastructure

Computing Resources:

• CPUs: General-purpose processors
• GPUs: Graphics processors excellent for parallel computations in deep learning
• TPUs: Tensor Processing Units specialized for AI workloads
• Cloud Platforms: AWS, Azure, Google Cloud providing scalable AI infrastructure

Development Tools:

• Frameworks: TensorFlow, PyTorch, scikit-learn
• Libraries: NumPy, Pandas for data processing
• MLOps Platforms: Tools for deploying and managing AI in production

Data Storage:

• Databases for structured data
• Data lakes for unstructured data
• Version control for datasets

The AI Lifecycle

Understanding the AI lifecycle is crucial for ISO 42001, which governs AI throughout its entire existence.

1. Planning and Design

Activities:

• Define business problem and objectives
• Assess feasibility of AI solution
• Consider alternatives (AI vs. non-AI approaches)
• Identify stakeholders and requirements
• Plan data needs and collection
• Design high-level system architecture
• Conduct initial risk assessment

AIMS Considerations:

• Align with organizational AI policy
• Involve ethics and governance review
• Document design decisions and rationale
• Establish success criteria and metrics

2. Data Collection and Preparation

Activities:

• Collect or acquire training data
• Clean and preprocess data
• Handle missing values and outliers
• Label data (for supervised learning)
• Split data into training/validation/test sets
• Explore and analyze data characteristics
• Address data quality issues

AIMS Considerations:

• Ensure data privacy and legal compliance
• Assess data for bias and representativeness
• Document data provenance and lineage
• Implement data governance controls
• Obtain necessary consents and rights

3. Model Development

Activities:

• Select model architecture and algorithms
• Configure hyperparameters
• Train model on training data
• Validate and tune using validation data
• Iterate to improve performance
• Experiment with different approaches
• Document experiments and results

AIMS Considerations:

• Apply responsible AI principles
• Test for bias and fairness
• Ensure model explainability where required
• Version control models and code
• Document model decisions and trade-offs

4. Evaluation and Validation

Activities:

• Test model on held-out test data
• Measure performance metrics
• Assess model across different subgroups
• Evaluate against business objectives
• Compare with baseline or alternative models
• Conduct stress testing and edge case analysis
• Validate with domain experts

AIMS Considerations:

• Test for fairness across protected groups
• Evaluate safety and robustness
• Assess compliance with requirements
• Document validation results
• Obtain stakeholder approval before deployment

5. Deployment

Activities:

• Prepare production environment
• Integrate model into systems and workflows
• Configure monitoring and logging
• Conduct user acceptance testing
• Train users and operators
• Establish human oversight mechanisms
• Create user documentation
• Plan rollout strategy (pilot, phased, full)

AIMS Considerations:

• Implement operational controls
• Ensure transparency to users
• Establish incident response procedures
• Configure performance monitoring
• Document deployment configuration

6. Operation and Monitoring

Activities:

• Process inference requests
• Monitor system performance
• Track accuracy and other metrics
• Collect feedback from users
• Log decisions for auditability
• Detect anomalies and issues
• Respond to incidents

AIMS Considerations:

• Continuous performance monitoring
• Bias and fairness monitoring
• Security monitoring
• User feedback collection
• Regular reporting to governance bodies

7. Maintenance and Updates

Activities:

• Address performance degradation
• Retrain models with new data
• Update models for changing conditions
• Fix bugs and issues
• Improve system based on feedback
• Version and release management
• Regression testing

AIMS Considerations:

• Change management processes
• Risk assessment for updates
• Validation of updated models
• Communication of changes to users
• Documentation updates

8. Decommissioning

Activities:

• Plan system retirement
• Migrate users to alternative solutions
• Archive data and models
• Delete sensitive data appropriately
• Document lessons learned
• Transfer knowledge

AIMS Considerations:

• Compliance with data retention requirements
• Secure disposal of data and models
• Stakeholder communication
• Knowledge preservation
• Final risk assessment

Lifecycle Nature: AI lifecycle is often iterative, with frequent cycles of monitoring, evaluation, and refinement rather than linear progression.

Key AI Performance Metrics

Classification Metrics

Accuracy: Proportion of correct predictions overall.

Formula: (True Positives + True Negatives) / Total Predictions

Precision: Of positive predictions, how many were actually positive?

Formula: True Positives / (True Positives + False Positives)

Use: When false positives are costly

Recall (Sensitivity): Of actual positives, how many were correctly identified?

Formula: True Positives / (True Positives + False Negatives)

Use: When false negatives are costly

F1 Score: Harmonic mean of precision and recall.

Formula: 2 × (Precision × Recall) / (Precision + Recall)

Use: Balancing precision and recall

Confusion Matrix: Table showing true positives, true negatives, false positives, and false negatives.

Use: Detailed view of classification performance

Regression Metrics

Mean Absolute Error (MAE): Average absolute difference between predictions and actual values.

Mean Squared Error (MSE): Average squared difference between predictions and actual values.

Root Mean Squared Error (RMSE): Square root of MSE, in same units as target variable.

R² Score: Proportion of variance in target variable explained by model.

Fairness Metrics

Demographic Parity: Equal selection rates across groups.

Equal Opportunity: Equal true positive rates across groups.

Equalized Odds: Equal true positive and false positive rates across groups.

Predictive Parity: Equal precision across groups.

Disparate Impact: Ratio of selection rates between groups.

Individual Fairness: Similar individuals receive similar outcomes.

Important AI Concepts for AIMS

Bias in AI

Definition: Systematic errors in AI outputs that create unfair outcomes for particular groups.

Sources:

• Historical Bias: Training data reflects past discrimination
• Sampling Bias: Training data doesn't represent population
• Measurement Bias: Proxy variables encode protected attributes
• Aggregation Bias: Model doesn't account for group differences
• Evaluation Bias: Benchmarks don't reflect real-world diversity

Mitigation:

• Use representative, diverse training data
• Test for bias across demographic groups
• Apply fairness constraints during training
• Implement bias detection and monitoring
• Involve diverse stakeholders in development

Explainability and Interpretability

Interpretability: Degree to which humans can understand model's decision-making process.

Explainability: Ability to provide human-understandable explanations for specific predictions.

Importance:

• Builds trust with users and stakeholders
• Enables debugging and improvement
• Supports accountability and compliance
• Required for high-stakes decisions
• Facilitates human oversight

Techniques:

• Feature importance analysis
• LIME (Local Interpretable Model-agnostic Explanations)
• SHAP (SHapley Additive exPlanations)
• Attention visualization
• Counterfactual explanations
• Model-agnostic methods

Trade-offs: More complex models (deep learning) often perform better but are harder to explain than simpler models (linear regression, decision trees).

Model Drift

Concept Drift: Statistical properties of target variable change over time.

Example: Customer preferences shifting, making old purchasing patterns less predictive

Data Drift: Distribution of input data changes over time.

Example: New types of transactions appearing after model deployment

Impact: Model performance degrades as real-world conditions diverge from training data.

Detection:

• Monitor prediction distribution
• Track model accuracy over time
• Compare current data to training data distribution
• Set up alerts for significant changes

Mitigation:

• Regular retraining with new data
• Online learning (continuous model updates)
• Ensemble methods combining multiple models
• Periodic model replacement

Overfitting and Underfitting

Overfitting: Model learns training data too well, including noise, failing to generalize to new data.

Symptoms: Excellent training performance, poor test performance

Causes: Model too complex, insufficient training data, training too long

Solutions: Regularization, cross-validation, more training data, early stopping

Underfitting: Model is too simple to capture underlying patterns.

Symptoms: Poor performance on both training and test data

Causes: Model too simple, insufficient features, inadequate training

Solutions: More complex model, better features, longer training

Balance: Good AI models find sweet spot between overfitting and underfitting, generalizing well to new data.

Robustness and Adversarial Examples

Robustness: AI system's ability to maintain performance under various conditions, including unexpected inputs and adversarial attacks.

Adversarial Examples: Inputs deliberately crafted to fool AI systems.

Example: Image with imperceptible perturbations causing misclassification

Adversarial Attacks:

• Evasion: Modifying inputs to avoid detection
• Poisoning: Injecting malicious data during training
• Model Extraction: Stealing model through queries
• Model Inversion: Reconstructing training data from model

Defenses:

• Adversarial training (training on adversarial examples)
• Input validation and sanitization
• Ensemble methods
• Detection of adversarial inputs
• Regular security testing

AI System Types Relevant to ISO 42001

Autonomous Systems

Characteristics: Make decisions and take actions with minimal human intervention.

Examples: Autonomous vehicles, robotic process automation, trading bots

AIMS Focus: Safety, reliability, human oversight, fail-safe mechanisms

Decision Support Systems

Characteristics: Provide recommendations to assist human decision-making.

Examples: Medical diagnosis assistance, loan application scoring, hiring candidate screening

AIMS Focus: Transparency, explainability, human-AI collaboration

Generative Systems

Characteristics: Create new content or data.

Examples: Large language models, image generators, code generators

AIMS Focus: Content authenticity, intellectual property, misinformation prevention

Perception Systems

Characteristics: Interpret sensory information from environment.

Examples: Facial recognition, speech recognition, object detection

AIMS Focus: Privacy, bias, accuracy, security

Optimization Systems

Characteristics: Find optimal solutions to complex problems.

Examples: Resource allocation, scheduling, routing

AIMS Focus: Fairness in optimization, transparency of constraints and objectives

Terminology Summary

Summary and Key Takeaways

Foundation for AIMS: Understanding AI concepts is essential for effective AI governance under ISO 42001.

AI Diversity: AI encompasses diverse technologies (ML, DL, NLP, computer vision, generative AI) requiring adaptable governance.

Lifecycle Approach: AI systems have distinct lifecycle phases from planning through decommissioning, each requiring governance.

Key Challenges: Bias, explainability, robustness, and drift are fundamental AI challenges addressed by AIMS.

Context Matters: Different AI system types (autonomous, decision support, generative) require different governance approaches.

Continuous Learning: AI field evolves rapidly; governance must adapt to new technologies and challenges.

Metrics and Measurement: Understanding performance metrics is crucial for AI evaluation and monitoring.

Data Centrality: Data quality and governance are foundational to AI performance and fairness.

Not Just Technical: AI governance requires understanding both technical concepts and their organizational, ethical, and regulatory implications.

Next Lesson: The AIMS Framework - Deep dive into ISO 42001's structure, clauses, and how they work together to create comprehensive AI governance.