AI and Machine Learning Definitions

AI Alignment

The field concerned with ensuring that artificial intelligence systems act in accordance with human values and intentions. AI alignment research focuses on developing methods to ensure AI systems pursue goals that are beneficial, safe, and aligned with human preferences, even as these systems become more capable and autonomous.

AI Governance

The frameworks, policies, and practices for ensuring the responsible development, deployment, and use of AI systems. AI governance encompasses legal regulations, ethical guidelines, industry standards, and organizational policies that guide how AI is created and implemented to ensure beneficial outcomes while minimizing risks and harms.

Algorithm vs. Model

An algorithm is a specific set of instructions or procedures for solving a problem or performing a computation. In machine learning, algorithms (like gradient descent) are the procedures used to train models.

A model is the result of applying an algorithm to data. It represents patterns learned from data and is used to make predictions or decisions. For example, a decision tree algorithm produces a decision tree model after training on data.

Artificial Intelligence (AI)

A broad field of computer science focused on creating systems that can perform tasks typically requiring human intelligence. While machine learning is a subset of AI, not all AI uses machine learning. Traditional AI systems might use rule-based approaches or expert systems that don't learn from data but instead follow pre-programmed instructions to make decisions.

Auditability

The ability to inspect, review, and verify how an AI system works, how it was developed, and how it produces specific outputs. Auditability enables third parties to examine an AI system's components, including its data sources, development processes, and decision-making patterns, to ensure compliance with standards, regulations, and ethical guidelines.

Bias in AI

Systematic errors in AI systems that can lead to unfair or discriminatory outcomes for certain groups or individuals. Bias can originate from various sources, including:

  • Training data that reflects historical or societal biases
  • Feature selection that overlooks important variables or includes proxy variables for protected attributes
  • Algorithm design choices that inadvertently amplify disparities
  • Deployment contexts that create feedback loops reinforcing existing biases

Addressing bias requires ongoing monitoring, diverse perspectives in AI development, and specific techniques to measure and mitigate unfair outcomes.

Black Box Machine Learning Algorithms

Algorithms whose internal workings and decision-making processes are not transparent or easily understood by humans. Despite potentially high accuracy, these models offer limited insight into how they reach specific conclusions.

Examples:

  • Deep neural networks with many layers
  • Random forests
  • Complex ensemble methods like gradient boosting machines
  • Support vector machines with non-linear kernels

Classification

A type of supervised machine learning where the algorithm learns to categorize data into predefined classes or categories. The model identifies patterns in labeled training data to make predictions about new, unseen data.

Examples: Email spam detection, medical diagnosis, customer churn prediction

Clustering

A type of unsupervised machine learning where the algorithm identifies natural groupings within data without predefined categories. The model groups similar data points together based on their inherent characteristics.

Examples: Customer segmentation, document grouping, anomaly detection

Computer Vision

A field of AI that enables computers to derive meaningful information from digital images, videos, and other visual inputs. Computer vision systems can identify objects, recognize faces, understand scenes, track movement, and even interpret human behaviors. It combines elements of machine learning, deep learning, and image processing to give machines the ability to "see" and understand visual information.

Counterfactual Explanations

A method of explaining AI decisions by showing how input factors would need to change to alter the outcome or prediction. These explanations answer the question "What would need to be different about this instance to get a different result?" For example, a loan application counterfactual might state: "If your annual income was $5,000 higher, your loan would have been approved." Counterfactuals are particularly valuable for providing actionable insights to individuals affected by AI decisions.

Data Drift/Model Drift

The phenomenon where the statistical properties of the data an AI system encounters during operation differ from those of the training data, causing model performance to degrade over time. Data drift occurs when real-world conditions change, making the model's learned patterns less accurate or relevant. Examples include changing customer preferences, new product features, or evolving language usage. Monitoring for drift and regularly retraining models are essential practices for maintaining AI system reliability.

Data Science

An interdisciplinary field that uses scientific methods, processes, algorithms, and systems to extract knowledge and insights from structured and unstructured data. Data science combines expertise from statistics, mathematics, computer science, and domain knowledge to analyze and interpret complex data and solve practical problems.

Deep Learning

A specialized subset of machine learning that uses neural networks with multiple layers (deep neural networks) to model complex patterns in data. Deep learning has revolutionized fields like computer vision, natural language processing, and speech recognition by automatically learning representations from raw data without requiring manual feature engineering. Common deep learning architectures include convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers.

Explainability

The degree to which an AI system's decisions and behaviors can be explained in human-understandable terms. Explainability focuses on providing clear, comprehensible explanations for why a system made a particular decision or prediction. This is crucial for applications in regulated industries, high-risk domains, and situations requiring human oversight.

Fairness Metrics

Quantitative measures used to assess whether an AI system treats different groups or individuals equitably. These metrics help identify and address biases in AI systems. Common fairness metrics include:

  • Demographic parity: Ensuring equal prediction rates across different groups
  • Equal opportunity: Ensuring equal true positive rates across groups
  • Predictive parity: Ensuring equal precision across groups
  • Individual fairness: Ensuring similar individuals receive similar predictions

Different fairness metrics may be appropriate in different contexts, and they sometimes conflict with each other, requiring careful consideration of the specific application and values at stake.

Feature/Feature Engineering

A feature is an individual measurable property or characteristic of a phenomenon being observed. In machine learning, features are the input variables used to make predictions.

Feature engineering is the process of selecting, transforming, or creating features to improve machine learning model performance. This may involve:

  • Selecting the most relevant variables
  • Creating new features from existing ones (e.g., calculating ratios or differences)
  • Encoding categorical variables
  • Scaling or normalizing numerical features
  • Handling missing values

Effective feature engineering often requires domain expertise and can significantly impact model performance.

Feature Importance

A measure of how much each feature in a model contributes to its predictions. Feature importance helps identify which variables most strongly influence the model's decisions, providing insights into the underlying patterns the model has learned. Different algorithms calculate feature importance in different ways:

  • Tree-based models measure how much each feature decreases impurity
  • Linear models use coefficient magnitudes
  • Permutation importance measures how model performance changes when a feature's values are randomly shuffled

Feature importance is a key tool for model interpretation and can guide feature selection and engineering.

Fine-tuning

The process of taking a pre-trained model and further training it on a specific dataset to adapt it to a particular task or domain. Fine-tuning leverages the general knowledge captured in a model trained on a large dataset and specializes it for a more targeted application. This approach is common with large language models and other deep learning systems, allowing organizations to benefit from models trained on vast datasets while customizing them for specific use cases with much less data and computing resources.

Foundation Models

Large AI models trained on vast amounts of data that can be adapted to a wide range of downstream tasks. Foundation models (like large language models and vision-language models) are characterized by their scale, generality, and emergent capabilities. Rather than being built for a single task, these models capture broad patterns that can be applied to many different applications through fine-tuning or prompting. Their influence has grown significantly due to their ability to reduce the need for task-specific data collection and model development.

Generative AI

AI systems that can create new content, including text, images, audio, video, code, and 3D models. Generative AI learns patterns from existing data and uses this knowledge to generate novel outputs that reflect those patterns while creating something new that wasn't explicitly in the training data. Recent advances in this field include large language models, diffusion models for image generation, and multimodal systems that can work across different types of content.

Glass Box (Explainable) Machine Learning Algorithms

Algorithms whose decision-making processes are transparent and can be understood by humans. These models allow users to trace exactly how inputs are transformed into outputs, making them valuable in regulated industries and sensitive applications.

Examples:

  • Decision trees and random forests (when limited in depth/complexity)
  • Linear and logistic regression
  • Rule-based systems
  • Bayesian models
  • k-nearest neighbors

Hyperparameters

Configuration settings that govern the training process of machine learning algorithms but are not learned from the data. Unlike model parameters (such as weights in a neural network), hyperparameters must be set before training begins. Examples include:

  • Learning rate in gradient descent
  • Number of trees in a random forest
  • Regularization strength in regression models
  • Number of hidden layers and neurons in a neural network
  • Batch size and number of epochs in deep learning

Finding optimal hyperparameters often requires techniques like grid search, random search, or Bayesian optimization.

Interpretability

The ability to understand the relationship between an AI system's inputs and outputs. A model is interpretable if a person can predict how changes to its input will affect its prediction. Interpretability focuses on how well humans can understand the model's mechanics rather than just explanations of individual decisions.

Large Language Model (LLM)

A type of AI system trained on vast amounts of text data to understand and generate human-like text. These neural network-based models can perform a wide range of language tasks such as writing, summarization, translation, and answering questions. LLMs learn the statistical patterns in language from their training data, enabling them to generate coherent and contextually relevant text.

LIME (Local Interpretable Model-agnostic Explanations)

A technique for explaining individual predictions of any machine learning model by approximating it locally with an interpretable model. LIME works by perturbing the input data and observing how the predictions change, then fitting a simple model (like a linear regression) around the specific prediction being explained. This creates a local explanation that identifies which features most influenced that particular decision, even for complex black-box models.

Machine Learning

A field of artificial intelligence that focuses on creating systems that can learn from and make predictions based on data. Instead of following explicitly programmed instructions, these systems identify patterns in data to improve their performance on a specific task over time. Machine learning enables computers to adapt and improve their performance without being explicitly programmed for every possible scenario.

Model Cards

Structured documentation that accompanies machine learning models to provide transparent information about their development, capabilities, limitations, and ethical considerations. Modeled after nutrition labels or medication information sheets, model cards typically include:

  • Intended uses and users
  • Training data characteristics
  • Evaluation results across different populations
  • Performance limitations and potential biases
  • Recommended and non-recommended applications
  • Environmental and social impact considerations

Model cards promote transparency and help stakeholders make informed decisions about model use.

Model Evaluation Metrics

Quantitative measures used to assess how well a machine learning model performs. Different metrics are appropriate for different tasks:

Classification metrics:

  • Accuracy: Proportion of correct predictions (both positive and negative)
  • Precision: Proportion of positive predictions that are correct
  • Recall/Sensitivity: Proportion of actual positives correctly identified
  • F1 Score: Harmonic mean of precision and recall
  • AUC-ROC: Area under the Receiver Operating Characteristic curve

Regression metrics:

  • Mean Absolute Error (MAE): Average of absolute differences between predictions and actual values
  • Mean Squared Error (MSE): Average of squared differences
  • Root Mean Squared Error (RMSE): Square root of MSE
  • R-squared: Proportion of variance explained by the model

Choosing appropriate evaluation metrics depends on the specific business problem and the relative costs of different types of errors.

Natural Language Processing (NLP)

A field at the intersection of AI, linguistics, and computer science that focuses on enabling computers to understand, interpret, and generate human language. NLP techniques power applications such as:

  • Machine translation
  • Sentiment analysis
  • Question answering systems
  • Text summarization
  • Speech recognition
  • Chatbots and virtual assistants

Modern NLP systems typically use deep learning approaches, including transformer-based architectures like those found in large language models.

Neural Network

A machine learning model inspired by the human brain's structure. Neural networks consist of interconnected nodes (neurons) organized in layers that process information. Each connection has a weight that adjusts during learning. Information flows through the network, with each neuron performing calculations and passing results to the next layer. Neural networks are particularly effective for complex tasks like image recognition, speech processing, and natural language understanding.

Overfitting and Underfitting

Overfitting occurs when a model learns the training data too well, including its noise and outliers, resulting in poor performance on new, unseen data. An overfit model captures random fluctuations rather than the underlying pattern.

Underfitting occurs when a model is too simple to capture the underlying pattern in the data, resulting in poor performance on both training and new data.

Finding the right balance between these extremes is crucial for creating models that generalize well to new data. Techniques to address these issues include:

  • Cross-validation
  • Regularization
  • Early stopping
  • Ensemble methods
  • Collecting more training data (for underfitting)

Post-hoc Explainability

Methods applied after a model has been trained to understand and explain its decisions, without changing the model itself. Post-hoc techniques are particularly valuable for complex black-box models like deep neural networks. Examples include:

  • Feature importance analysis
  • LIME and SHAP for local explanations
  • Partial dependence plots to show relationships between features and predictions
  • Activation visualization in neural networks
  • Concept activation vectors to identify high-level concepts

These approaches help stakeholders understand model behavior without sacrificing the performance advantages of complex models.

Prompt Engineering

The practice of designing, refining, and optimizing prompts (input instructions) to effectively elicit desired behaviors from large language models and other generative AI systems. Prompt engineering involves crafting inputs that guide the model toward producing specific types of outputs, following particular formats, or adhering to certain constraints. Skilled prompt engineering can dramatically improve the quality, relevance, and safety of AI-generated content without changing the underlying model.

Recommendation Systems

AI systems that suggest items or content to users based on their preferences, behavior, or characteristics. These systems power features like "customers who bought this also bought," content feeds, playlist suggestions, and personalized marketing. Common approaches include:

  • Collaborative filtering: Recommending items based on similar users' preferences
  • Content-based filtering: Recommending items similar to what a user has liked before
  • Hybrid approaches: Combining multiple recommendation strategies
  • Knowledge-based systems: Using explicit knowledge about users and items

Recommendation systems balance accuracy, diversity, novelty, and serendipity to create valuable user experiences.

Regression

A type of supervised machine learning where the algorithm learns to predict continuous numerical values (rather than categories). The model analyzes the relationship between independent variables (features) and a dependent variable (target) to make predictions.

Examples: House price prediction, sales forecasting, temperature prediction

Reinforcement Learning

A machine learning approach where an agent learns to make decisions by taking actions in an environment to maximize some notion of cumulative reward. Unlike supervised learning, reinforcement learning doesn't rely on labeled examples but instead on feedback from the environment. The agent learns through trial and error, balancing exploration of unknown options with exploitation of known successful strategies. Applications include game playing, robotics, autonomous vehicles, resource management, and recommendation systems.

Responsible AI

A framework for developing and deploying AI systems that are ethical, fair, transparent, and accountable. Responsible AI practices aim to ensure that AI technologies benefit humanity while minimizing potential harms. Key principles include:

  • Fairness and non-discrimination
  • Privacy and security
  • Transparency and explainability
  • Human oversight and control
  • Accountability for outcomes
  • Robustness and safety
  • Environmental and social sustainability

Implementing responsible AI requires a multidisciplinary approach, involving technical, legal, ethical, and business considerations.

SHAP (SHapley Additive exPlanations)

A method for explaining individual predictions from any machine learning model based on game theory concepts. SHAP assigns each feature an importance value for a particular prediction by considering all possible combinations of features and measuring how each feature contributes to the prediction. SHAP values have desirable mathematical properties that make them particularly reliable for consistent explanations, including local accuracy, missingness, and consistency. They can be visualized in various ways to provide insights into both individual predictions and overall model behavior.

Structured Data

Data organized in a predefined format, typically stored in databases or spreadsheets with a clear schema. Structured data includes information like dates, names, addresses, and numerical values that fit neatly into rows and columns. This type of data is easily searchable and analyzable using traditional methods and query languages like SQL.

Supervised Machine Learning

A machine learning approach where the algorithm is trained on labeled data, meaning each training example is paired with the correct answer or outcome. The algorithm learns to map inputs to outputs based on these examples, enabling it to make predictions on new, unseen data. The model's performance can be objectively evaluated by comparing its predictions to the known correct answers.

Training, Testing, and Validation Data

Training data is the portion of data used to teach a machine learning model by adjusting its parameters based on patterns in this data.

Validation data is used during the training process to tune hyperparameters and evaluate model performance on data not seen during training, helping prevent overfitting.

Testing data is completely separate from both training and validation data and is used only after the model is finalized to estimate how well it will perform on new, real-world data.

Properly splitting data into these sets is crucial for creating models that generalize well. Common approaches include:

  • Simple random splits (e.g., 70% training, 15% validation, 15% testing)
  • Cross-validation for smaller datasets
  • Time-based splits for temporal data
  • Stratified sampling to maintain class distributions

Transfer Learning

A machine learning technique where knowledge gained from training a model on one task is applied to a different but related task. Transfer learning leverages the features and patterns learned from a source domain (often with abundant data) to improve performance on a target domain (often with limited data). This approach can significantly reduce the amount of data and computational resources needed to develop effective models for specific applications. Common approaches include:

  • Using pre-trained models as feature extractors
  • Fine-tuning pre-trained models on new data
  • Domain adaptation to bridge differences between source and target domains

Transparency

Openness about how AI systems are designed, developed, and deployed. Transparency includes clear documentation about data sources, model architecture, training methods, and performance metrics. It enables stakeholders to understand what the system does, how it operates, and what its limitations are. Transparency is foundational for responsible AI deployment and building user trust.

Unstructured Data

Data that lacks a predefined format or organization and doesn't fit neatly into traditional databases. Examples include text documents, emails, social media posts, images, videos, and audio files. This type of data requires specialized techniques for processing and analysis, such as natural language processing for text or computer vision for images.

Unsupervised Machine Learning

A machine learning approach where the algorithm learns patterns and relationships from unlabeled data without explicit guidance. The system must discover the inherent structure of the data on its own. Common applications include clustering (grouping similar items), dimensionality reduction (simplifying complex data while preserving important information), and anomaly detection (identifying unusual patterns).