The terms in this glossary are divided into two groups: Core terms and Advanced / Specialized terms.
The core terms are those you’re most likely to meet in articles, products, interviews, and day-to-day conversations about AI. The advanced/specialized terms go into more detail on particular techniques, technologies, or subfields.
When we say “core”, we refer to the basic ideas that:
- Show up again and again, in various AI contexts;
- Make it easier to understand what people are talking about in presentations, blog posts, and product documentation.
Advanced/specialised terminology embraces sets of terms such as:
- Specific techniques, model architectures, and variations;
- Governance and safety concepts;
- Niche ideas that mostly appear in particular areas; for example, knowledge graphs in knowledge management, or NeRF in 3D vision.
Core AI Terms
Big-picture AI concepts
Artificial Intelligence (AI)
AI is when computers are built to do things that usually need human intelligence, like understanding language, recognizing images, or making decisions. It’s about making machines “smart” in specific tasks.
Machine Learning (ML)
ML is a way for computers to learn from data instead of being explicitly programmed with every rule. The more examples you give it, the better it usually gets at a task.
Deep Learning
Deep learning is a type of machine learning that uses many-layered neural networks, loosely inspired by the brain. It’s behind things like voice assistants, image recognition, and advanced chatbots.
Generative AI (GenAI)
Generative AI creates new content—like text, images, code, or music—based on patterns it learned from data. It doesn’t just classify or detect; it actually produces something new.
Generative Pre-trained Transformer (GPT)
GPT is a type of generative AI model trained on huge amounts of text so it can predict and generate the next words in a sentence. That’s what lets it write emails, answer questions, or chat like a person.
Large Language Model (LLM)
An LLM is an AI model trained on massive text datasets so it can understand and generate human-like language. “Large” refers to both the size of the model (many parameters) and the amount of data it learned from.
Transformer
A Transformer is a kind of model architecture that excels at understanding sequences like sentences using an attention mechanism. It’s the core design behind modern LLMs like GPT.
Conversational AI
Conversational AI is technology that allows computers to talk with people in natural language, by voice or text. It powers things like virtual assistants, chatbots, and voicebots.
AI Chatbot
A chatbot is a program that talks with users through messages, often to answer questions or help with tasks. An AI chatbot uses machine learning to give more natural, flexible, and context-aware replies.
Computer Vision
Computer vision is AI that lets computers “see” and understand images and videos. It’s used for things like face recognition, object detection, and reading text from pictures.
Cognitive Computing
Cognitive computing aims to mimic human-like thinking, such as understanding language, reasoning, and learning from experience. It supports humans in making decisions rather than just following fixed rules.
Enterprise AI
Enterprise AI is the use of AI inside organizations to improve business tasks—like forecasting, customer support, or automation. It focuses on reliability, security, and scaling across a company.
AI Automation
AI automation uses AI to handle tasks that were done by people, such as data entry, document processing, or simple decision-making. The goal is to make work faster, cheaper, and less error-prone.
AI Workforce / AI Worker
An AI worker is a software “agent” that can perform tasks like a digital coworker—answering emails, summarizing documents, or updating systems. The “AI workforce” is the collection of these agents working alongside humans.
Data & learning basics
Dataset
A dataset is a collection of data, like a big table or bundle of examples, that an AI model learns from or is tested on. Think of it as the “study material” for AI.
Training Data
Training data is a part of the dataset used to teach the model how to make predictions. The model looks at these examples over and over to learn patterns.
Training Set
The training set is simply the chunk of data specifically chosen for training the model. It’s where the model “practices” before being tested.
Validation Data / Test Set
Validation or test data is used after training to check how well the model learned, using examples it hasn’t seen before. It helps detect overfitting and measure real performance.
Structured Data
Structured data is neatly organized in fixed formats, like rows and columns in a spreadsheet (e.g., age, price, date). It’s easy for both humans and machines to work with.
Semi-structured Data
Semi-structured data has some organization but not a strict table format—for example JSON, XML, or log files. There are recognizable tags or fields, but it’s more flexible than a spreadsheet.
Unstructured Data
Unstructured data has no fixed format, like plain text, images, audio, or videos. It’s rich and messy, and AI is often used to make sense of it.
Data Preprocessing / Pre-processing
Data preprocessing is cleaning and preparing raw data so it’s usable for a model—like removing errors, filling gaps, or converting text to numbers. It’s the “washing and chopping” step before cooking.
Data Augmentation
Data augmentation means creating extra training examples by slightly changing existing ones, like flipping an image or rephrasing a sentence. This helps the model generalize better without collecting new real-world data.
Data Labeling
Data labeling is adding the correct answers or tags to data, like marking an email as “spam” or “not spam.” These labels tell the model what the right output should be during training.
Data Extraction
Data extraction is pulling relevant information out of a source, such as reading names from documents or values from PDFs. It turns raw content into usable structured data.
Data Ingestion
Data ingestion is the process of bringing data from different sources into a storage or processing system. It’s like piping water from many taps into one big tank.
Data Discovery
Data discovery is finding out what data you have, where it lives, and what it contains. It helps teams understand which data is useful for analysis or AI projects.
Data Drift / Model Drift
Data drift or model drift happens when the real-world data changes over time, so the model’s old learning no longer fits well. As a result, its predictions slowly get worse unless you retrain or update it.
Core ML types & algorithms
Supervised Learning
Supervised learning is when a model learns from examples that already have the correct answers (labels), like pictures labeled “cat” or “dog.” It tries to learn a mapping from input to the known output.
Unsupervised Learning
Unsupervised learning is when a model looks at data without labels and tries to find patterns on its own, like grouping similar customers together. It discovers structure instead of learning from right/wrong answers.
Reinforcement Learning
Reinforcement learning is when an AI learns by trial and error, getting rewards or penalties as feedback. It’s like training a pet: good actions are rewarded so they happen more often.
Self-supervised learning
Self-supervised learning lets a model create its own training labels from raw data, like hiding part of a sentence and asking the model to predict the missing words. It’s a clever way to learn from huge unlabeled datasets.
Transfer Learning
Transfer learning means taking a model that’s already trained on one task and adapting it to a new, related task. It’s like learning Spanish faster because you already know Italian.
Fine-tuning
Fine-tuning is a type of transfer learning where you start with a pre-trained model and train it a bit more on your specific data. This lets you customize a powerful model without training from scratch.
Model Tuning / Hyperparameter Tuning / Hyperparameters
Hyperparameters are settings that control how a model learns (like learning rate or number of layers), not learned from data directly. Hyperparameter tuning is the process of trying different settings to get the best performance.
Model Training
Model training is the process where the AI adjusts its internal settings by looking at training data and reducing errors step by step. It’s the “learning” phase before the model is used for real predictions.
Model
A model is the learned mathematical “brain” that takes input data and produces an output, like predicting a price or generating text. It’s the end result of training.
Parameters / Model Parameter
Parameters are the internal numbers inside a model (like weights in a neural network) that get adjusted during training. Together, they define how the model transforms input into output.
Objective Function / Loss Function
The loss function measures how wrong the model’s predictions are compared to the correct answers. The goal of training is to minimize this loss so the model gets better.
Gradient Descent
Gradient descent is a method for updating model parameters step by step in the direction that reduces the loss. Think of it like walking downhill towards the lowest point in a landscape of errors.
Random Forest
Random Forest is a machine learning method that combines many decision trees and averages their results to make better, more stable predictions. It’s like asking a “forest” of mini-experts and taking a consensus.
Recurrent Neural Networks (RNN)
RNNs are neural networks designed to handle sequences, like sentences or time series, by “remembering” previous steps. They pass information from one time step to the next.
Convolutional Neural Networks (CNN / CNNs)
CNNs are neural networks specialized for grid-like data such as images, using filters that slide over the input to detect patterns like edges or shapes. They’re widely used in image recognition.
Artificial Neural Network (ANN)
An ANN is a network of connected nodes (“neurons”) inspired by the brain, which transforms inputs through layers to produce an output. Many modern AI methods, like deep learning, are built from ANNs.
Evaluation & performance
Accuracy
Accuracy is the percentage of predictions the model got right overall. It’s “how often was the model correct?”
Precision
Precision measures how many of the items the model flagged as positive were actually positive. It answers: “When the model says ‘yes’, how often is it right?”
Recall
Recall measures how many of the actual positive items the model successfully found. It answers: “Out of all the real ‘yes’ cases, how many did the model catch?”
F-score (F-measure, F1 measure)
The F1 score combines precision and recall into one number by taking their harmonic mean. It’s useful when you want a balance between catching positives and being correct when you say “positive.”
Overfitting
Overfitting is when a model memorizes the training data too closely and doesn’t generalize well to new data. It performs great on seen examples but poorly on unseen ones.
Underfitting
Underfitting is when a model is too simple to capture the patterns in the data. It performs badly even on the training data because it hasn’t learned enough.
Regularization
Regularization is a set of techniques used to stop overfitting by gently penalizing overly complex models. It encourages the model to learn simpler, more general patterns.
Model Evaluation
Model evaluation is the process of checking how good a trained model is using metrics like accuracy, precision, and recall on validation or test data. It tells you whether the model is ready for real use.
Model Monitoring
Model monitoring is watching a model’s performance over time in the real world to catch problems like data drift or accuracy drops. It ensures the model stays reliable after deployment.
NLP & text-focused fundamentals
Natural Language Processing (NLP)
NLP is the field of AI that helps computers work with human language, like reading, writing, and understanding text. It powers things like chatbots, translation, and spam filters.
Natural Language Understanding (NLU)
NLU is the part of NLP focused on understanding the meaning and intent behind text or speech. It helps systems figure out what a user is really asking or saying.
Natural Language Generation (NLG)
NLG is about getting computers to produce human-like text. It turns data or ideas into sentences, paragraphs, or full documents.
Natural Language Query (NLQ)
NLQ means asking questions in normal human language instead of using a special query language. For example, typing “Show me sales from last month” instead of writing a SQL query.
Natural Language Technology (NLT)
NLT is a broad term for all technologies that handle human language, covering NLP, NLU, NLG, and related tools. It’s basically the toolbox for working with text and speech.
Token
A token is a small unit of text that a language model processes, often a word or piece of a word. Models “see” text as sequences of tokens, not letters or full sentences directly.
Token Limit
The token limit is the maximum number of tokens a model can handle in one go (input + output). If you exceed it, the model has to cut off or ignore part of the text.
Context Window
The context window is how much text (in tokens) the model can “remember” at once for a single request. Anything outside this window is basically forgotten by the model during that interaction.
Annotation
Annotation is adding labels or notes to text, such as marking entities, sentiment, or parts of speech. These labels help train and evaluate NLP models.
Part-of-Speech (POS) Tagging / Part-of-Speech Tagging
POS tagging means labeling each word in a sentence with its role, like noun, verb, or adjective. It helps the model understand sentence structure.
Named Entity Recognition (NER)
NER finds and labels important items in text, like people, companies, locations, and dates. For example, marking “Paris” as a location in a sentence.
Sentiment Analysis
Sentiment analysis detects the emotion or opinion in text, such as positive, negative, or neutral. It’s often used to analyze customer reviews or social media posts.
Text Analytics / Text Mining / Computational Linguistics
These terms refer to methods for analyzing large amounts of text to find patterns, trends, or insights. It can include things like keyword extraction, topic detection, and frequency counts.
Text Summarization
Text summarization creates a shorter version of a document that keeps the key information. It can be done by picking important sentences or by generating new condensed text.
Machine Translation
Machine translation automatically converts text or speech from one language to another, like English to Spanish. Modern systems often use deep learning and large language models.
Question & Answer (Q&A)
Q&A systems take a user’s question in natural language and try to return a direct, relevant answer. Think of search that gives you an answer, not just a list of links.
Speech Recognition
Speech recognition turns spoken language into written text. It’s what powers voice typing and systems like “Hey Siri” or “OK Google.”
Speech Analytics
Speech analytics analyzes audio conversations (like call center recordings) to find patterns, topics, emotions, or problems. It helps businesses understand what customers say and how agents respond.
Working with models in practice
Inference
Inference is when a trained model is used to make predictions or generate outputs on new data. It’s the “using the model” phase, not the “learning” phase.
Inference Engine
An inference engine is the part of a system that runs the trained model to produce answers, often optimized for speed and efficiency. It’s like the “runtime” that serves predictions.
Model Deployment
Model deployment is putting a trained model into a real environment (like a web app or internal tool) so people or systems can use it. It’s the step where the model leaves the lab and goes into production.
Model Versioning
Model versioning means keeping track of different versions of a model over time, like v1, v2, v3. It helps you know which model is running, roll back if needed, and compare performance.
Model Compression
Model compression is shrinking a model so it uses less memory and runs faster, often with little loss in accuracy. This is especially useful for phones, browsers, or small devices.
Foundation Model
A foundation model is a large, general-purpose model trained on huge amounts of data that can be adapted to many tasks. GPT-style models are examples: they’re a “base” you build on.
Custom Model / Custom/Domain Language Model
A custom model is an AI model adapted or built for a specific domain or company need, like finance documents or medical notes. It understands your particular data and use cases better than a generic model.
Edge model
An edge model is a model that runs on devices at the “edge” (like phones, cameras, or IoT devices) instead of in the cloud. This can make responses faster and more private.
Pre-training
Pre-training is the initial phase where a model learns general patterns from a huge, broad dataset. It gives the model a wide base of knowledge before it’s specialized.
Post-training
Post-training refers to additional steps after the main training, such as fine-tuning, alignment, or optimization. It’s where you refine the model for specific tasks or make it safer and more useful.
Pretrained model
A pretrained model is a model that has already gone through pre-training and is ready to be fine-tuned or used as-is. It lets you start from something smart instead of from scratch.
Prompting & LLM-centric ideas
Prompt
A prompt is the text or instructions you give an AI model to tell it what you want. It’s your “question” or “task description.”
AI Prompt
An AI prompt is simply a prompt written specifically for an AI system, often carefully phrased to get the best answer. It can include instructions, examples, and constraints.
Prompt Engineering
Prompt engineering is the craft of designing and wording prompts so the AI gives better, more accurate, or more useful responses. It’s like learning how to “talk” to the model effectively.
Prompt chaining
Prompt chaining means breaking a big task into smaller steps and using multiple prompts in sequence, passing the output of one into the next. It helps solve complex problems in a more controlled way.
Completions
A completion is the text that the AI model generates to continue or answer your prompt. You give the start (prompt); the model produces the rest (completion).
Context Window
The context window is how much text (in tokens) the model can “see” and use at once for a single request. Anything beyond that limit can’t be considered in its answer.
Retrieval Augmented Generation (RAG)
RAG is a technique where the model first fetches relevant documents or facts from a database, then uses them to generate an answer. It helps the AI stay more accurate and up-to-date instead of relying only on what it was trained on.
Data structures & knowledge representation
Knowledge Graph
A knowledge graph is a network of facts represented as connected nodes (things) and edges (relationships), like “Alice —works at→ Company X.” It helps machines understand how pieces of information relate to each other.
Knowledge Model
A knowledge model is the structured way you represent what a system “knows” about a domain—its concepts, relationships, and rules. It’s like a blueprint of knowledge for an AI or software system.
Ontology
An ontology is a formal description of the things in a domain (like “Person,” “Company,” “Product”) and how they relate. It’s a detailed, agreed-upon vocabulary plus rules for how terms connect.
Taxonomy
A taxonomy is a hierarchical classification, like a tree of categories and subcategories (e.g., “Animal → Mammal → Dog”). It organizes items into “is-a” relationships to make them easier to browse and manage.
Thesauri
A thesaurus (in this context) is a structured list of terms that shows synonyms, related terms, and sometimes broader/narrower ones. It helps keep language consistent while allowing for variations in wording.
Simple Knowledge Organization System (SKOS)
SKOS is a standard way to represent things like taxonomies and thesauri so different systems can share and understand them. Think of it as a common format for structured vocabularies on the web.
Controlled Vocabulary
A controlled vocabulary is a fixed, approved list of terms that everyone must use for tagging or classifying content. It reduces confusion by avoiding multiple words for the same thing.
Trust, ethics & governance – conceptually core
AI Ethics
AI ethics is about the principles and values that should guide how AI is built and used, like fairness, transparency, and respect for people’s rights. It asks, “Is this AI right to do, not just possible?”
Responsible AI / Responsible AI by Design
Responsible AI means designing, building, and using AI in a way that is safe, fair, and accountable from the very beginning. “By design” means these concerns are built in, not added at the end.
Ethical AI
Ethical AI is AI that follows agreed moral guidelines, such as avoiding harm, discrimination, or deception. It’s the goal that AI systems behave in ways we consider morally acceptable.
AI Bias
AI bias happens when an AI system treats some people or groups unfairly, often because its training data or design reflects real-world biases. This can lead to skewed results, like worse outcomes for certain demographics.
Algorithmic Fairness / AI Fairness
Algorithmic fairness is about measuring and reducing unfair differences in how an AI system treats different groups. The aim is for the AI to make decisions that are as fair as possible across people.
Explainable AI (XAI) / Explainable AI / Explainability
Explainable AI focuses on making AI decisions understandable to humans—why did it make that prediction? It helps people trust and check the system instead of treating it as a black box.
AI Governance
AI governance is the set of policies, processes, and controls that guide how AI is developed and used within an organization or society. It’s like “company rules and oversight” for AI.
AI Guardrails
AI guardrails are limits and safety rules built into AI systems to prevent harmful or unwanted behavior. For example, blocking certain topics or enforcing strict data-use rules.
AI Safety
AI safety is about making sure AI systems don’t cause harm—whether accidentally or intentionally—both now and as they become more capable. It includes preventing misuse, accidents, and unexpected behaviors.
Human-in-the-Loop (HITL)
Human-in-the-loop means a human is actively involved in key steps of an AI system, like reviewing decisions or giving feedback. It keeps people in control rather than leaving everything to the machine.
Human-Centered AI
Human-centered AI is designed around human needs, values, and limitations, supporting people instead of replacing or ignoring them. It focuses on usability, well-being, and real-world impact.
AI Privacy
AI privacy is about protecting people’s personal data when AI systems are trained or used. It aims to ensure sensitive information isn’t exposed, misused, or collected without proper consent.
Advanced/Specialized AI Terms
“Future of AI” & philosophical
Artificial General Intelligence (AGI)
AGI is a hypothetical AI that can understand, learn, and perform any intellectual task a human can. It’s not just good at one job, but flexible and general like a person.
Artificial Super Intelligence (ASI)
ASI is a theoretical AI that is smarter than humans in basically every way—science, creativity, social skills, and more. It’s beyond human-level intelligence.
Singularity
The singularity is the idea that at some point AI could improve itself so quickly that society changes in a sudden, dramatic way. It’s often linked to AGI or ASI and big, unpredictable shifts.
Chinese Room
The Chinese Room is a thought experiment arguing that a machine could appear to understand language (by following rules) without truly understanding it. It questions whether AI can ever be genuinely conscious or just “pretend smart.”
Turing Test
The Turing Test checks whether a machine’s conversation is indistinguishable from a human’s. If a human judge can’t tell which is which, the machine is said to have “passed” the test.
Advanced LLM & generative techniques
Chain of Thought
Chain of thought is when an AI shows or uses step-by-step reasoning instead of jumping straight to an answer. It’s like letting the model “think out loud” to solve complex problems.
Prompt Injection
Prompt injection is when a user sneaks instructions into input (like a document or message) that override the system’s original rules. It’s a kind of attack to make the model behave in unintended ways.
Prompt Leaking
Prompt leaking is when an AI accidentally reveals its hidden instructions or system messages. It exposes how the model is set up, which can be a security or safety issue.
Frequency Penalty
A frequency penalty is a setting that makes the model less likely to repeat the same words or phrases. It helps reduce boring or repetitive responses.
Presence Penalty
A presence penalty discourages the model from using words it has already mentioned, encouraging it to bring in new topics. It makes responses more diverse and exploratory.
Top-k Sampling
Top-k sampling tells the model to choose the next word only from the top k most likely options. It limits choices to the “best few” to balance creativity and coherence.
Top-p (Nucleus Sampling)
Top-p sampling lets the model pick from the smallest set of words whose combined probability is at least p (like 0.9). It adapts the choice size based on how confident the model is.
Stop Sequences
Stop sequences are special strings of text that tell the model, “Stop generating now.” They help cut off answers cleanly at the right moment.
Instruction Tuning
Instruction tuning is training a model on many examples of “instruction → good response” so it follows natural-language instructions better. It makes the model more obedient and helpful.
Few-Shot Learning
Few-shot learning is when you show the model just a few examples in the prompt and it learns the pattern on the fly. It’s like giving 2–3 demos and saying “Now do the same.”
Zero-Shot Learning / Zero shot extraction
Zero-shot learning is when the model handles a new task without being given any explicit examples in the prompt. You simply describe the task in words and it figures it out.
Multitask prompt tuning (MPT)
Multitask prompt tuning is adjusting a model using small prompt-like parameters so it can handle many tasks efficiently. It’s a lightweight way to specialize a big model without retraining everything.
Mixture of Experts
Mixture of Experts is a model design where many “expert” sub-models exist, and only a few are activated for each input. It can make large models more efficient and specialized.
Multimodal AI / Multimodal models and modalities
Multimodal AI can work with more than one type of data at once, like text + images or audio + video. It can, for example, look at a picture and read a caption to answer questions.
Image Generation / Image-to-Image Translation
Image generation creates new images from text or noise, like “draw a cat on the moon.” Image-to-image translation transforms one image into another style or form, like turning a sketch into a realistic picture.
Image Captioning
Image captioning takes an image and automatically generates a description in text. It’s like an AI saying, “This picture shows a dog playing in a park.”
Image Colorization
Image colorization adds color automatically to black-and-white images. It’s often used to restore old photos or stylize images.
Image Denoising
Image denoising removes visual noise or grain from images to make them cleaner and clearer. It’s like using a smart “cleanup” filter.
Image Inpainting
Image inpainting fills in missing or damaged parts of an image in a realistic way. For example, it can remove an object and fill the gap with matching background.
Image Segmentation
Image segmentation separates an image into regions or objects, like marking which pixels belong to “person,” “car,” or “background.” It helps systems understand the structure of the scene.
Image Upscaling
Image upscaling increases the resolution of an image, often using AI to add plausible details instead of just stretching pixels. It makes small or blurry images sharper and larger.
NeRF (Neural Radiance Fields)
NeRF is a technique that uses a neural network to represent a 3D scene from 2D images, allowing you to render new viewpoints. It’s like building a 3D model from a handful of photos.
Style Transfer
Style transfer takes the style of one image (like a painting) and applies it to another image (like a photo). You might end up with a photo that looks like it was painted by Van Gogh.
Generative Summarization
Generative summarization creates a summary by writing new sentences that capture the main ideas of a text. It doesn’t just copy; it paraphrases and condenses.
Extractive Summarization
Extractive summarization builds a summary by selecting and stitching together the most important sentences from the original text. It copies key pieces instead of generating new wording.
Architectures, libraries & named models
BERT (Bidirectional Encoder Representations from Transformers)
BERT is a language model that reads text both left-to-right and right-to-left to better understand context. It’s widely used for tasks like search, Q&A, and classification.
RoBERTa
RoBERTa is an improved version of BERT trained longer and on more data with some tweaks. It usually performs better than the original BERT on many language tasks.
XLNet
XLNet is a language model that learns by predicting words in many different orders, not just left-to-right. This helps it capture richer relationships in text.
CLIP (Contrastive Language-Image Pre-training)
CLIP is a model that learns from images paired with text so it can connect what it “sees” with what it “reads.” It can, for example, match a caption to the right image.
TensorFlow
TensorFlow is a popular open-source library for building and training machine learning and deep learning models. It gives developers tools to define and run complex AI pipelines.
TPU
A TPU (Tensor Processing Unit) is a special chip made by Google to speed up AI calculations, especially for TensorFlow models. It’s optimized for the kinds of math deep learning uses.
GPU
A GPU (Graphics Processing Unit) is a chip originally designed for graphics but now heavily used to train AI models. It can perform many simple calculations in parallel, which is perfect for deep learning.
More technical ML / training concepts
Back Propagation / Forward Propagation
Forward propagation is when data flows through the model to produce a prediction. Backpropagation is when the model calculates how wrong it was and adjusts its parameters to do better next time.
Hidden Layer
A hidden layer is a layer of neurons in a neural network between the input and output. It transforms the data step by step to learn complex patterns.
Latent Space
Latent space is the model’s internal “concept space” where data is represented in compressed, abstract form. Similar things in the real world tend to end up close together in this space.
Embedding
An embedding is a vector (a list of numbers) that represents something like a word, sentence, or image in a way the model can work with. Similar items have similar embeddings.
Objective Function (in depth)
The objective (or loss) function is a formula that tells the model how bad its predictions are. Training tries to minimize this value so the model improves.
Double Descent
Double descent is a phenomenon where, as models get larger, their error first goes down, then up, then down again. It challenges the old idea that “bigger models after a point always overfit.”
End-to-End Learning
End-to-end learning means training one big model to go directly from raw input to final output, instead of using many separate hand-designed steps. The model learns to handle the whole pipeline itself.
Incremental Learning
Incremental learning lets a model update itself gradually as new data arrives, without retraining from scratch. It’s useful when data is streaming or changes over time.
Federated Learning
Federated learning trains a model across many devices (like phones) without moving the raw data to a central server. Only model updates are shared, which can help protect privacy.
Self-supervised learning (advanced side of it)
In self-supervised learning, the model creates training tasks from the data itself (like predicting missing parts) so it can learn useful representations. It’s a powerful way to use huge unlabeled datasets.
Regularization (mathematical details)
Regularization adds extra terms to the loss function or applies constraints to discourage overly complex models. Mathematically, it nudges parameters toward simpler values to reduce overfitting.
Model Drift / Data Drift
Data drift is when the input data changes over time (for example, customer behavior shifts). Model drift is the resulting decline in model performance because it was trained on older patterns.
Model Explainability / Model Interpretability (technical methods)
These are techniques that show how a model made a decision, such as highlighting important features or generating simplified surrogate models. They help engineers and users inspect and trust the model’s behavior.
Advanced NLP & linguistic concepts
Disambiguation
Disambiguation is the process of figuring out which meaning of a word or phrase is intended, like whether “bank” means a riverbank or a financial institution. It’s key for understanding language correctly.
Morphological Analysis
Morphological analysis breaks words into their smallest meaningful parts, like roots and endings (“played” → “play” + “-ed”). It helps models understand word forms and variations.
Parsing
Parsing analyzes a sentence’s structure, showing how words group together and relate (like subject, verb, object). It gives the model a “grammar tree” to work with.
Cataphora
Cataphora is when a word refers to something that appears later in the sentence or text, like “Before she spoke, Anna took a breath.” The reference points forward instead of backward.
SAO (Subject-Action-Object) / Triple or Triplet Relations
SAO breaks information into triples like “Customer – buys – product.” These simple structures make it easier for systems to capture who did what to whom.
Semantic Network
A semantic network is a graph where nodes are concepts and edges are relationships, similar to a knowledge graph but focused on meaning. It visually and structurally represents how ideas connect.
Semantics
Semantics is about the meaning of words, phrases, and sentences, not just their form. In AI, it means understanding what text means, not just how it looks.
Similarity (and Correlation)
Similarity measures how alike two items are (often using embeddings), while correlation measures how two numeric variables change together. Both help detect relationships in data.
Windowing
Windowing means looking at a fixed-size slice (“window”) of a longer sequence, like a few words in a row or a time segment. It helps models handle long data by processing it chunk by chunk.
Co-occurrence
Co-occurrence is when two words or items appear together frequently in data, like “coffee” and “cup.” Models use this to infer that these things are related.
Corpus / Specialized corpora
A corpus is a large collection of text used to train or test language models. A specialized corpus focuses on a specific domain, like medical articles or legal documents.
Knowledge & search systems
Semantic Search
Semantic search looks for results based on meaning rather than exact keywords. It tries to understand what you meant, not just what you typed.
Cognitive Map
A cognitive map is a mental model of how concepts relate to each other. In AI, it’s a structured representation that helps systems reason about knowledge.
Linked Data
Linked Data is a way of publishing data so that different datasets connect through shared identifiers. It lets machines follow “links” between pieces of information across the web.
Knowledge Engineering
Knowledge engineering is the process of capturing expert knowledge and putting it into structures (rules, ontologies, graphs) that computers can use. It’s like turning human expertise into machine-readable form.
Insight Engines
Insight engines are advanced search tools that combine AI, NLP, and sometimes graphs to find and surface relevant insights from scattered data. They go beyond keyword search to “understand” content and context.
Intelligent Document Processing (IDP) / IDEP
IDP uses AI to read, classify, and extract data from documents like invoices or contracts. It turns messy PDFs and scans into structured, usable information.
Auto-classification
Auto-classification is when AI automatically assigns documents or items to categories or tags. It saves humans from manually sorting large volumes of content.
Categorization / Category Trees
Categorization is sorting items into groups (categories) like “Invoices,” “Contracts,” or “Support Tickets.” Category trees arrange these groups in a hierarchy, from broad to specific.
Content Enrichment
Content enrichment adds extra metadata or links to raw content, like tagging people, places, or topics. It makes content easier to search, filter, and analyze.
Entity / ETL
An entity is a specific thing in text, like a person, company, or location. ETL here refers to recognizing and extracting these entities from documents so they can be stored and analyzed.
Governance, risk, and enterprise-y AI
AI Risk Management
AI risk management is finding, assessing, and reducing risks from AI systems, like bias, security issues, or failures. It’s about making AI safer and more reliable in real use.
AI Auditing
AI auditing means systematically checking how an AI system behaves, what data it uses, and whether it follows rules and policies. It’s like an internal or external “inspection” of the AI.
AI Orchestration
AI orchestration coordinates multiple AI models, tools, and workflows so they work together smoothly. Think of it as the conductor managing many “AI instruments” in a pipeline.
AI Automation vs Traditional AI
Traditional automation follows fixed, rule-based scripts, while AI automation uses learning models that can handle fuzzier, less predictable tasks. AI automation is better at unstructured data and changing conditions.
AI Guardrails (implementation patterns)
AI guardrails are technical controls like filters, safety checks, and policies that limit what an AI can say or do. Implementation patterns include input validation, output filtering, and policy-based blocking.
Environmental, Social, and Governance (ESG) in AI context
ESG in AI means considering environmental impact, social fairness, and governance standards when developing and deploying AI. It frames AI as part of a company’s broader responsibility to society.
Workplace 5.0
Workplace 5.0 is a vision of work where humans and intelligent systems collaborate closely. AI handles repetitive or complex digital tasks while humans focus on creativity, judgment, and relationships.
Enterprise AI (beyond basic definition)
Enterprise AI is not just “AI in companies”; it includes robust governance, security, scalability, integration with business systems, and clear ROI. It’s AI built to meet the messy realities of large organizations.
More specialized evaluation & production concepts
Hallucinate/Hallucination / Hallucinations (detection & mitigation)
A hallucination is when an AI confidently makes up false information. Detection and mitigation involve checking AI outputs against trusted data, adding retrieval (like RAG), and using guardrails to reduce wrong answers.
Grounding
Grounding means tying an AI’s answers to real, verifiable data sources instead of letting it rely only on its training. It helps make responses more factual and trustworthy.
Responsible AI (as frameworks and processes)
Responsible AI frameworks define policies, roles, and steps for building and using AI safely and fairly. They turn high-level ethics ideas into concrete processes, checklists, and governance.
Data Scarcity
Data scarcity is when you don’t have enough good-quality data to train a model well. It often leads to poor performance or overfitting.
Data Augmentation strategies
Data augmentation strategies create extra training examples by modifying existing ones (e.g., rotating images, paraphrasing text). This helps models generalize better, especially when data is scarce.
Model Monitoring (drift metrics, alerts)
Model monitoring tracks metrics like accuracy and drift over time, raising alerts if performance drops or data changes. It’s how you catch problems early and decide when to retrain.
Niche but interesting
Deepfake
A deepfake is synthetic media (video, audio, or image) generated by AI that convincingly imitates a real person. It can be used for both creative applications and harmful misinformation.
Big Data
Big Data refers to datasets that are so large, fast, or complex that traditional tools struggle to handle them. AI and distributed systems are often used to process and analyze this data.
Agents & Agentic AI / Agentic Process Automation (APA)
Agents and agentic AI are systems that can plan, take actions, and interact with tools to achieve goals with some autonomy. APA applies such agents to business processes, automating multi-step workflows end to end.
AI Alignment
AI alignment is about making sure an AI’s goals and behavior match human values and intentions. It’s trying to ensure “what the AI optimizes for” is truly what we want.
AI Assistant
An AI assistant is a digital helper that can answer questions, perform tasks, and integrate with tools. It acts like a smart coworker or personal assistant in software form.
Pattern Recognition
Pattern recognition is the ability of AI to detect regularities in data—like shapes in images or trends in numbers. It’s a core skill behind most machine learning tasks.
Predictive Analytics
Predictive analytics uses data and models to estimate what is likely to happen in the future, like predicting churn or demand. It turns historical data into forward-looking insights.
Actionable Intelligence
Actionable intelligence is information that is not just interesting but directly useful for making decisions or taking action. It answers “so what?” in a way that can drive concrete steps.
Content
In this context, content usually means digital information such as documents, emails, web pages, images, or videos. It’s the raw material that AI and search systems analyze.
Treemap
A treemap is a visual chart that shows hierarchical data as nested rectangles, where size (and sometimes color) encodes value. It’s handy for seeing how parts make up a whole at a glance.
Plugins
Plugins are add-on components that extend the capabilities of a system, like giving an AI access to a calendar, database, or third-party app. They let you connect the model to real tools and data.
Tunable
Tunable means a model or system has settings (hyperparameters, options, knobs) you can adjust to fit your needs. The more tunable something is, the more you can optimize it for your specific use case.