Introduction to Artificial Intelligence
Expert-defined terms from the Undergraduate Certificate in AI for Public Policy and Governance course at LearnUNI. Free to read, free to share, paired with a professional course.
Artificial Intelligence – The field of computer science focused on creati… #
Related terms: machine learning, reasoning, perception. Explanation: AI encompasses algorithms that enable computers to learn, reason, plan, perceive, and interact. Example: An AI chatbot that answers citizen queries about public services. Application: Policy analysis tools that simulate economic impacts of legislation. Challenge: Ensuring transparency and avoiding bias in decision‑making processes.
Algorithm – A step‑by‑step computational procedure for solving a problem… #
Related terms: complexity, heuristic, optimization. Explanation: Algorithms define the logic that AI models follow to process data. Example: A sorting algorithm that organizes voter registration records. Application: Efficient data processing for large‑scale public datasets. Challenge: Selecting algorithms that balance accuracy with computational cost.
Artificial Neural Network – A computational model inspired by the structu… #
Related terms: deep learning, backpropagation, activation function. Explanation: Consists of layers of interconnected nodes that transform input data into predictions. Example: A network predicting traffic congestion for urban planning. Application: Forecasting demand for public transportation. Challenge: Interpreting the “black‑box” nature of deep models for policy stakeholders.
Bias (Algorithmic) – Systematic error that skews outcomes in favor of or… #
Related terms: fairness, discrimination, data quality. Explanation: Bias can arise from training data, model design, or deployment context. Example: An AI system that underestimates loan eligibility for minority applicants. Application: Auditing AI tools used in welfare allocation. Challenge: Detecting hidden biases and implementing mitigation strategies.
Big Data – Extremely large and complex datasets that traditional processi… #
Related terms: volume, velocity, variety, veracity. Explanation: Big data provides the raw material for training AI models. Example: Real‑time sensor streams from smart city infrastructure. Application: Analyzing patterns of energy consumption to guide policy. Challenge: Managing privacy concerns and ensuring data security.
Classification – A supervised learning task that assigns input data to pr… #
Related terms: label, training set, confusion matrix. Explanation: Models learn from labeled examples to predict class membership for new data. Example: Classifying social media posts as supportive or critical of a policy. Application: Sentiment analysis for public opinion monitoring. Challenge: Dealing with imbalanced class distributions and ambiguous cases.
Clustering – An unsupervised learning technique that groups similar data… #
Related terms: k‑means, hierarchical, silhouette score. Explanation: Algorithms discover natural structures within datasets. Example: Grouping neighborhoods based on demographic and mobility patterns. Application: Targeted outreach for community engagement programs. Challenge: Determining the appropriate number of clusters and interpreting results.
Computer Vision – The field enabling machines to interpret and process vi… #
Related terms: object detection, image segmentation, convolutional neural network. Explanation: Uses AI techniques to extract meaning from visual data. Example: Analyzing satellite imagery to detect illegal land use. Application: Monitoring environmental compliance in protected areas. Challenge: Handling varied lighting conditions and ensuring ethical use of surveillance.
Cross‑Validation – A statistical method for assessing how a predictive mo… #
Related terms: train‑test split, k‑fold, overfitting. Explanation: Data is partitioned into subsets; models are trained on some and validated on others. Example: Using 5‑fold cross‑validation to evaluate a crime‑prediction model. Application: Selecting robust models for policy impact assessments. Challenge: Balancing computational load with the need for reliable performance estimates.
Data Governance – The overall management of data availability, usability,… #
Related terms: metadata, stewardship, compliance. Explanation: Establishes policies and procedures for handling data throughout its lifecycle. Example: A municipal data portal that defines access rights for public datasets. Application: Ensuring reliable data pipelines for AI‑driven decision support. Challenge: Coordinating across agencies and aligning with privacy regulations.
Data Preprocessing – The set of operations applied to raw data to prepare… #
Related terms: cleaning, normalization, feature engineering. Explanation: Involves handling missing values, scaling, encoding categorical variables, etc. Example: Converting timestamps to standardized formats for temporal analysis. Application: Improving the accuracy of predictive models used in budget forecasting. Challenge: Maintaining reproducibility and avoiding inadvertent data distortion.
Decision Tree – A supervised learning model that splits data based on fea… #
Related terms: entropy, Gini impurity, pruning. Explanation: Represents decisions as a tree structure, facilitating interpretability. Example: A tree that predicts eligibility for social assistance based on income and household size. Application: Transparent policy rule generation for automated eligibility checks. Challenge: Preventing overfitting and managing high‑dimensional feature spaces.
Deep Learning – A subset of machine learning that uses multilayer neural… #
Related terms: representation learning, GPU acceleration, autoencoder. Explanation: Deep architectures can automatically learn hierarchical features from raw data. Example: A deep model that forecasts air‑quality indices from sensor networks. Application: Real‑time environmental monitoring for health policy interventions. Challenge: High computational demands and difficulty in explaining predictions to non‑technical stakeholders.
Ethical AI – The practice of designing, developing, and deploying AI syst… #
Related terms: responsibility, accountability, transparency. Explanation: Emphasizes fairness, privacy, inclusivity, and avoidance of harm. Example: An AI tool that recommends public housing placements while ensuring equitable access. Application: Guiding the creation of AI governance frameworks for government agencies. Challenge: Translating abstract ethical concepts into concrete technical specifications.
Explainable AI (XAI) – Techniques that make the behavior of AI models und… #
Related terms: interpretability, model‑agnostic, SHAP, LIME. Explanation: Provides insights into why a model made a particular decision. Example: Visualizing feature contributions for a predictive policing algorithm. Application: Building trust with citizens when AI influences public services. Challenge: Balancing explanation depth with model performance.
Feature Engineering – The process of creating informative variables from… #
Related terms: feature selection, transformation, domain knowledge. Explanation: Involves deriving new attributes, encoding, and scaling. Example: Constructing a “distance to nearest clinic” feature for health‑access models. Application: Enhancing predictive accuracy of disease‑outbreak detection systems. Challenge: Requires domain expertise and can be time‑consuming.
Fuzzy Logic – A form of reasoning that handles approximate rather than ex… #
Related terms: membership function, inference, rule base. Explanation: Allows AI systems to work with uncertainty and vagueness. Example: Controlling traffic lights based on fuzzy assessments of congestion. Application: Adaptive policy mechanisms that respond to ambiguous indicators. Challenge: Designing appropriate membership functions and ensuring consistency.
Generative Adversarial Network (GAN) – A deep learning architecture consi… #
Related terms: synthetic data, adversarial training, latent space. Explanation: The generator creates data samples, while the discriminator evaluates their authenticity. Example: Generating synthetic demographic profiles to augment scarce survey data. Application: Protecting privacy while enabling AI model training on public datasets. Challenge: Controlling mode collapse and preventing misuse of generated content.
Gradient Descent – An optimization algorithm that iteratively adjusts mod… #
Related terms: learning rate, convergence, stochastic. Explanation: Computes the gradient of the loss and moves parameters in the opposite direction. Example: Training a linear regression model to predict unemployment rates. Application: Fine‑tuning policy‑impact models for better accuracy. Challenge: Selecting appropriate learning rates and avoiding local minima.
Graph Neural Network (GNN) – A neural architecture designed to operate on… #
Related terms: nodes, edges, message passing. Explanation: Captures relationships and dependencies between entities. Example: Modeling inter‑agency collaboration networks to identify bottlenecks. Application: Analyzing social networks for misinformation spread. Challenge: Scaling to large, dynamic graphs and interpreting learned representations.
Heuristic – A rule‑of‑thumb or practical method used to solve problems mo… #
Related terms: approximation, greedy algorithm, search. Explanation: Provides feasible solutions with reduced computational effort. Example: A heuristic that selects the most vulnerable neighborhoods for resource allocation. Application: Rapid scenario planning during emergencies. Challenge: May produce suboptimal outcomes and require validation.
Human‑in‑the‑Loop (HITL) – An approach where human judgment complements a… #
Related terms: oversight, feedback, augmentation. Explanation: Humans review, correct, or guide AI outputs before final decisions. Example: Analysts verify AI‑generated risk scores for public health interventions. Application: Ensuring accountability in AI‑driven policy enforcement. Challenge: Designing efficient interfaces and preventing human bias from re‑entering the loop.
Hyperparameter – Configuration settings external to the model that influe… #
Related terms: grid search, tuning, regularization. Explanation: Includes learning rate, batch size, number of layers, etc. Example: Setting the depth of a decision tree to avoid overfitting. Application: Optimizing AI models for budget‑impact analysis. Challenge: Searching large hyperparameter spaces can be computationally expensive.
Inference – The process of applying a trained AI model to new, unseen dat… #
Related terms: deployment, latency, batch processing. Explanation: Distinct from training; focuses on real‑time or offline prediction. Example: Using a trained model to infer traffic patterns for a city’s congestion pricing policy. Application: Delivering decision support dashboards for policymakers. Challenge: Ensuring scalability and maintaining model performance over time.
Instance Segmentation – A computer‑vision task that classifies each pixel… #
Related terms: mask R‑CNN, bounding box, pixel‑wise labeling. Explanation: Provides fine‑grained object detection and delineation. Example: Identifying each illegal dumping site in aerial photographs. Application: Enforcing environmental regulations through automated monitoring. Challenge: Requires large annotated datasets and high computational resources.
Internet of Things (IoT) – Network of physical devices embedded with sens… #
Related terms: edge computing, smart city, telemetry. Explanation: Generates continuous streams of data useful for AI analytics. Example: Sensors measuring water quality across a municipal supply network. Application: Real‑time policy adjustments for resource management. Challenge: Managing data heterogeneity and securing vulnerable devices.
Knowledge Graph – A structured representation of entities and their relat… #
Related terms: ontology, semantic web, SPARQL. Explanation: Enables AI systems to reason over linked data. Example: A graph linking legislation, agencies, and affected demographic groups. Application: Facilitating policy impact analysis through queryable relationships. Challenge: Keeping the graph up‑to‑date and handling ambiguous relationships.
Label (Training Data) – The ground‑truth annotation that indicates the co… #
Related terms: supervised learning, annotation, ground truth. Explanation: Essential for training models to recognize patterns. Example: Tagging survey responses as “support” or “oppose” a proposed tax. Application: Building sentiment classifiers for public consultation platforms. Challenge: Obtaining high‑quality labels at scale and mitigating annotator bias.
Latent Variable – An unobserved variable inferred from observed data, oft… #
Related terms: dimensionality reduction, factor analysis, hidden state. Explanation: Captures hidden factors that influence observable outcomes. Example: Latent variables representing socioeconomic status in a health risk model. Application: Enhancing predictive power while reducing dimensionality. Challenge: Interpreting latent dimensions and ensuring they align with policy concepts.
Linear Regression – A statistical method that models the relationship bet… #
Related terms: coefficients, residuals, R‑squared. Explanation: Provides a simple, interpretable baseline for prediction. Example: Predicting tax revenue based on economic indicators. Application: Quick policy impact estimates before more complex modeling. Challenge: Assumes linearity and may miss nonlinear interactions.
Machine Learning (ML) – A subset of AI that enables computers to learn pa… #
Related terms: supervised, unsupervised, reinforcement. Explanation: Algorithms improve performance as they are exposed to more data. Example: An ML model that predicts the likelihood of a natural disaster affecting a region. Application: Allocating emergency resources proactively. Challenge: Data quality, model drift, and interpretability in public‑sector contexts.
Model Drift – The degradation of model performance over time due to chang… #
Related terms: concept drift, monitoring, retraining. Explanation: Occurs when real‑world conditions evolve away from training data. Example: A crime‑prediction model that becomes less accurate after a new policing policy is enacted. Application: Setting up continuous monitoring pipelines for AI‑driven services. Challenge: Detecting drift early and updating models without service interruption.
Natural Language Processing (NLP) – The field that enables computers to u… #
Related terms: tokenization, sentiment analysis, transformer. Explanation: Uses statistical and neural techniques to process text and speech. Example: An NLP system that extracts policy‑relevant topics from legislative documents. Application: Automating the summarization of public comments on proposed regulations. Challenge: Managing multilingual data and preserving nuance in policy language.
Neural Architecture Search (NAS) – Automated methods for designing optima… #
Related terms: search space, reinforcement learning, efficiency. Explanation: Searches for architectures that achieve high performance with limited resources. Example: Discovering a compact model for on‑device air‑quality prediction. Application: Deploying AI in low‑resource government offices. Challenge: High computational cost and ensuring discovered architectures meet regulatory constraints.
Overfitting – When a model captures noise in the training data, resulting… #
Related terms: regularization, validation, complexity. Explanation: The model performs well on training data but fails on unseen examples. Example: A predictive model that memorizes specific past incidents but cannot forecast future trends. Application: Implementing regularization techniques to improve policy‑impact forecasts. Challenge: Detecting overfitting early and balancing model flexibility with robustness.
Precision (Metric) – The proportion of true positive predictions among al… #
Related terms: recall, F1 score, confusion matrix. Explanation: Measures exactness; high precision indicates few false positives. Example: In a fraud‑detection system for welfare payments, precision reflects how many flagged cases are actual fraud. Application: Prioritizing investigations where resources are limited. Challenge: Trade‑off with recall; optimizing for one may degrade the other.
Privacy‑Preserving Machine Learning – Techniques that protect individual… #
Related terms: differential privacy, federated learning, encryption. Explanation: Methods like adding noise or training locally across devices. Example: Federated learning across municipal health departments to predict disease spread without sharing raw patient records. Application: Collaborative AI initiatives across agencies with strict privacy mandates. Challenge: Maintaining model accuracy while adhering to legal privacy standards.
Reinforcement Learning (RL) – A learning paradigm where agents interact w… #
Related terms: policy, reward function, Markov decision process. Explanation: Agents learn optimal actions through trial and error. Example: An RL agent that allocates limited public housing units to maximize social welfare over time. Application: Dynamic resource allocation for disaster response. Challenge: Defining appropriate reward structures that align with policy goals.
Recall (Metric) – The proportion of true positive cases that the model co… #
Related terms: precision, sensitivity, F1 score. Explanation: Measures completeness; high recall indicates few false negatives. Example: In a health‑screening AI, recall shows how many at‑risk individuals are correctly flagged. Application: Ensuring vulnerable populations are not missed in social programs. Challenge: Balancing recall with precision to avoid overwhelming resources with false alarms.
Regularization – Techniques that add constraints to a model to prevent ov… #
Related terms: L1, L2, dropout, penalty. Explanation: Encourages simpler models that generalize better. Example: Applying L2 regularization to a logistic regression predicting unemployment benefits eligibility. Application: Building stable AI tools for long‑term policy monitoring. Challenge: Selecting appropriate regularization strength for specific datasets.
Responsible AI – A framework that ensures AI systems are designed, deploy… #
Related terms: governance, accountability, risk assessment. Explanation: Integrates principles of fairness, transparency, and inclusivity. Example: A responsible AI checklist for a predictive budgeting tool. Application: Institutionalizing AI oversight in governmental decision‑making. Challenge: Translating high‑level principles into actionable processes and metrics.
Risk Assessment (AI) – The systematic evaluation of potential harms and u… #
Related terms: impact analysis, mitigation, stakeholder analysis. Explanation: Identifies threats such as bias, privacy violations, and operational failures. Example: Assessing the risk of an AI system that automates welfare eligibility decisions. Application: Informing policy safeguards and regulatory compliance. Challenge: Quantifying intangible risks and incorporating diverse stakeholder perspectives.
Sampling Bias – Systematic error introduced when a sample does not repres… #
Related terms: selection bias, representativeness, stratification. Explanation: Leads to skewed model predictions and unfair outcomes. Example: Training a health‑prediction model on data from urban hospitals only, neglecting rural populations. Application: Adjusting data collection strategies for inclusive AI in public health. Challenge: Detecting hidden biases and correcting them without compromising data integrity.
Scalable AI – Designing AI systems that can handle growing data volumes a… #
Related terms: distributed computing, cloud services, parallelism. Explanation: Uses architectures that expand horizontally across machines. Example: Deploying a city‑wide traffic‑prediction model on a cloud platform that scales during peak hours. Application: Supporting nationwide policy simulations that require massive computation. Challenge: Managing cost, latency, and data sovereignty across jurisdictions.
Semantic Segmentation – Assigning a class label to every pixel in an imag… #
Related terms: U‑Net, pixel classification, encoder‑decoder. Explanation: Provides fine‑grained understanding of visual data. Example: Mapping flood‑affected areas from satellite images to guide relief efforts. Application: Real‑time disaster response mapping for policy makers. Challenge: High annotation effort and need for robust models under varying conditions.
Supervised Learning – A machine‑learning paradigm where models are traine… #
Related terms: classification, regression, loss function. Explanation: The algorithm learns to map inputs to correct outputs. Example: Predicting the success probability of grant applications based on past award data. Application: Automating evaluation processes in public funding agencies. Challenge: Obtaining sufficient high‑quality labeled data and preventing overreliance on historical patterns.
Support Vector Machine (SVM) – A classification algorithm that finds the… #
Related terms: kernel trick, margin, support vectors. Explanation: Effective for high‑dimensional spaces and small datasets. Example: Classifying policy documents into “environmental” or “economic” categories. Application: Rapid sorting of large legislative corpora for thematic analysis. Challenge: Choosing appropriate kernels and scaling to massive datasets.
Transfer Learning – Leveraging knowledge from a pre‑trained model on a re… #
Related terms: fine‑tuning, domain adaptation, pretraining. Explanation: Reduces data requirements and accelerates development. Example: Adapting a language model trained on general news to analyze public comments on a new tax law. Application: Deploying AI quickly for emerging policy issues. Challenge: Avoiding negative transfer when source and target domains differ significantly.
Uncertainty Quantification – Techniques that estimate the confidence or p… #
Related terms: Bayesian inference, confidence interval, predictive variance. Explanation: Provides decision makers with risk awareness. Example: Reporting confidence bounds for projected greenhouse‑gas emissions reductions. Application: Informed policy planning that accounts for prediction uncertainty. Challenge: Computing uncertainty efficiently for complex deep models.
Unsupervised Learning – Learning patterns from data without explicit labe… #
Related terms: clustering, dimensionality reduction, autoencoder. Explanation: Models infer relationships based solely on input features. Example: Detecting anomalous patterns in utility usage without predefined categories. Application: Early warning systems for fraud or abuse in public programs. Challenge: Interpreting results and validating that discovered patterns are meaningful for policy.
Variational Autoencoder (VAE) – A generative model that learns a probabil… #
Related terms: encoder, decoder, KL divergence. Explanation: Enables data generation and dimensionality reduction. Example: Synthesizing realistic demographic profiles to augment training data while preserving privacy. Application: Creating benchmark datasets for AI policy research. Challenge: Balancing reconstruction quality with latent space regularization.
Version Control (Data & Models) – Systematic tracking of changes to datas… #
Related terms: Git, reproducibility, provenance. Explanation: Facilitates collaboration, auditability, and rollback capabilities. Example: Maintaining a repository of model versions used for different fiscal year forecasts. Application: Ensuring transparent AI development pipelines in government agencies. Challenge: Managing large binary assets and integrating with existing bureaucratic workflows.
Virtual Assistant – An AI‑driven software agent that can understand natur… #
Related terms: chatbot, dialogue system, intent recognition. Explanation: Interacts with users via text or voice interfaces. Example: A virtual assistant answering citizen queries about tax filing deadlines. Application: Reducing administrative workload and improving public service accessibility. Challenge: Maintaining up‑to‑date knowledge bases and handling ambiguous requests.
Weighted Majority Algorithm – An online learning method that combines pre… #
Related terms: ensemble, regret, expert advice. Explanation: Adjusts influence of each predictor over time. Example: Combining forecasts from economic, health, and environmental models to produce a composite policy impact score. Application: Robust decision support that integrates diverse expert systems. Challenge: Selecting appropriate expert pools and preventing dominance of a single inaccurate predictor.
Zero‑Shot Learning – A technique that enables models to recognize classes… #
Related terms: semantic embedding, attribute transfer, generalization. Explanation: Uses descriptions or relationships to infer unseen categories. Example: Classifying a newly introduced policy area based on textual description without prior examples. Application: Rapid adaptation to emerging policy domains without extensive data collection. Challenge: Ensuring sufficient semantic information to achieve reliable predictions.