AI Dictionary
The Most Important AI Terms Explained Clearly
Cut through the AI noise with this simple, beginner-friendly dictionary of essential terms.
From “Deep learning” to “Hyperparameter,” this glossary makes AI easy to understand.
-
An intelligent entity, often a software program or a physical robot, that can perceive its environment through sensors (e.g., cameras, microphones, data feeds) and act upon that environment through actuators (e.g., motors, speakers, API calls) to achieve specific goals.
Example: Example: A self-driving car (the agent) uses cameras and lidar (sensors) to perceive its surroundings and controls the steering wheel, accelerator, and brakes (actuators) to reach a destination (goal). A chatbot (the agent) perceives user input (sensor) and responds with text (actuator) to answer questions or provide information (goal).
-
A precise set of instructions or rules designed to perform a specific task or solve a particular problem. In AI, algorithms are the computational engines that enable systems to learn, reason, and make decisions.
Example: Example: The gradient descent algorithm is used to train neural networks by iteratively adjusting the model's parameters to minimize errors. A classification algorithm like a decision tree uses a series of rules to categorize data into different classes (e.g., spam or not spam).
-
A theoretical form of AI possessing human-level cognitive abilities. An AGI would be capable of understanding, learning, and applying knowledge across a vast range of tasks, exhibiting consciousness, reasoning, and problem-solving skills comparable to a human. Currently, all existing AI systems are considered narrow AI or weak AI, excelling only in specific domains.
Example: Example (Hypothetical): An AGI could seamlessly switch between writing a novel, diagnosing a medical condition, composing music, and strategizing a complex business plan, demonstrating adaptability and understanding across diverse fields.
-
The overarching field of computer science dedicated to creating intelligent agents – systems that can reason, learn, perceive, understand language, and solve problems in a way that mimics human cognitive processes.
Example: AI encompasses a wide range of technologies, including machine learning algorithms that power recommendation systems (like Netflix suggesting movies), natural language processing that enables virtual assistants (like Siri or Alexa), and computer vision that allows facial recognition software to identify individuals.
-
A crucial component in many modern neural network architectures, particularly in natural language processing (NLP) and computer vision. It allows the model to dynamically weigh the importance of different parts of the input data when processing information and making predictions.
Example: When translating a sentence, an attention mechanism helps the model align words in the source language with their corresponding words in the target language. In image captioning, it allows the model to focus on specific regions of the image when generating the description.
-
A fundamental and widely used algorithm for training artificial neural networks. It works by calculating the gradient of the loss function (the error the model makes) with respect to each of the network's weights. This gradient information is then propagated backward through the network to update the weights and reduce the loss.
Example: If a neural network misclassifies an image of a cat, backpropagation calculates how each layer contributed to the error and adjusts the weights to improve future accuracy.
-
A field of AI focused on enabling machines to interpret and understand visual data from the world, such as images and videos.
Example: Used in self-driving cars to detect pedestrians and traffic signs, in medical imaging to spot anomalies, and in security for facial recognition.
-
A deep learning architecture ideal for image data. CNNs use filters in convolutional layers to automatically detect patterns like edges, textures, and objects.
Example: CNNs power image classification systems — identifying objects in images like cats, cars, or traffic signs.
-
A subfield of machine learning using neural networks with many layers to learn complex patterns in large datasets.
Example: Used in voice assistants, image recognition, autonomous driving, and language translation systems.
-
Techniques used to increase dataset size and diversity by modifying existing data (e.g., rotating, flipping, cropping, adding noise).
Example: Augmenting cat images with different angles and lighting to improve model robustness in recognizing cats in varied conditions.
-
A structured collection of data points used to train, validate, and evaluate machine learning models. Quality and diversity of the dataset directly affect model performance.
Example: A spam detection dataset may include email content, sender info, and a label like “spam” or “not spam.”
-
Techniques designed to make AI decisions more transparent and interpretable, especially important in high-stakes domains like healthcare or finance.
Example: A medical AI system might explain a diagnosis by showing which symptoms contributed most to the prediction.
-
Quantitative criteria used to measure a machine learning model's performance, often based on task type (classification, regression, etc.).
Example: In email spam detection, metrics like accuracy, precision, recall, and F1-score are used to judge model effectiveness.
-
The process of transforming raw data into meaningful features that improve model accuracy and interpretability.
Example: From customer data, deriving “days since last purchase” or “total spent” can be more predictive of churn than raw logs.
-
A transfer learning technique where a pre-trained model is adapted to a specific task by continuing training on a smaller, task-specific dataset.
Example: A language model trained on general text is fine-tuned using customer service transcripts to specialize in support queries.
-
A model made of two neural networks — a generator and a discriminator — that compete to create and evaluate realistic synthetic data.
Example: GANs generate photorealistic images of people, deepfakes, and synthetic art or music.
-
An optimization algorithm used to minimize a function — typically the model's loss — by updating parameters in the direction of steepest descent.
Example: Like walking downhill blindfolded by feeling the slope, gradient descent moves model weights to reduce prediction error.
-
A model configuration value set before training begins — such as learning rate, batch size, or number of layers — that controls the training process but is not learned from the data itself.
Example: The number of neurons per layer in a neural network or the learning rate used to update weights during training.
-
The process of using a trained model to make predictions or decisions based on new, unseen data.
Example: A spam detection model, after training, uses inference to classify each incoming email as spam or not.
-
A learning strategy where the model memorizes training data and makes predictions based on the similarity to stored instances, rather than learning general patterns.
Example: The k-nearest neighbors (KNN) algorithm compares a new data point to its nearest labeled neighbors to classify it.
-
The method by which AI systems organize, store, and access information to support reasoning, learning, and decision-making.
Example: Knowledge graphs represent concepts and their relationships (e.g., “Elon Musk” — “CEO of” — “Tesla”) to enable complex queries and inference.
-
A hyperparameter that controls how large a step the model takes when updating weights during training. It affects training speed and convergence quality.
Example: A high learning rate might skip over the optimal solution; a low one might make training very slow or get stuck in local minima.
-
A mathematical function used to measure how far off a model's predictions are from the actual outcomes. The goal of training is to minimize this loss.
Example: In regression, mean squared error is a common loss function; in classification, cross-entropy is frequently used.
-
A deep learning model trained on massive amounts of text to perform language-based tasks like writing, summarizing, translating, and answering questions.
Example: GPT-3, BERT, and Claude are LLMs that can generate human-like responses and perform complex language tasks.
-
A subset of AI focused on developing algorithms that allow computers to learn patterns from data and make decisions without being explicitly programmed.
Example: ML powers email spam filters, recommendation engines, fraud detection, and predictive text in smartphones.
-
The final result of training a machine learning algorithm on data. It maps inputs to outputs and encapsulates learned patterns.
Example: A trained image classifier that predicts whether a photo contains a dog or a cat is a model.
-
A field of AI focused on enabling machines to understand, interpret, generate, and manipulate human language (spoken or written).
Example: NLP powers voice assistants, language translation, sentiment analysis, and chatbots.
-
A computational model inspired by the human brain, made up of layers of interconnected "neurons" that process data and learn complex patterns.
Example: A neural network can classify images, predict stock prices, or analyze medical scans based on the data it has learned from.
-
A modeling issue where a machine learning model learns the training data too well — including noise and irrelevant patterns — and performs poorly on new data.
Example: A student memorizing exact test answers but failing on a similar test with new wording; likewise, an overfit model can't generalize.
-
A metric for evaluating classification models. It measures how many of the items predicted as positive are actually positive.
Example: In spam detection, high precision means most emails labeled “spam” truly are spam, minimizing false alarms.
-
A metric that measures how many actual positive cases were correctly identified by the model. It’s the model’s ability to “catch” all the true positives.
Example: In disease detection, high recall ensures sick patients aren’t missed, even if it means more false positives.
-
-
A type of neural network designed for sequential data. RNNs maintain a memory of previous inputs, enabling them to learn patterns over time.
Example: Used in language modeling to predict the next word in a sentence or in speech recognition to process audio frames in sequence.
-
A machine learning approach where models are trained on labeled data — input-output pairs — to learn how to predict outcomes for new inputs.
Example: Training a spam filter using emails labeled as “spam” or “not spam” so the model can learn to classify new emails accurately.
-
A supervised learning algorithm used for classification and regression. It tries to find the optimal boundary (hyperplane) between classes that maximizes the margin between them.
Example: Given two flower types, an SVM finds the best dividing line so it can classify new flowers accurately based on their features.
-
A deep learning architecture that relies entirely on attention mechanisms and is highly effective at handling sequential data, especially in NLP.
Example: Transformers power models like GPT and BERT. When translating a sentence, the model can consider all words simultaneously using attention.
-
The portion of data used to teach a machine learning model. It contains inputs and expected outputs so the model can learn patterns and relationships.
Example: For a dog breed classifier, training data includes labeled images of various breeds that the model learns from.
-
A technique where a model trained on one task is adapted to a new but related task, often requiring less data and training time for the second task.
Example: A model trained on general images can be fine-tuned to classify medical X-rays using a much smaller medical dataset.
-
A type of machine learning where the model learns patterns or structure from data without labeled outputs. It's often used for clustering or dimensionality reduction.
Example: Grouping customers into segments based on behavior without knowing their categories in advance.
-
A portion of data held out from training that's used to tune model parameters and prevent overfitting. It helps assess how well the model generalizes.
Example: During training, a drop in validation accuracy while training accuracy rises signals overfitting.
-
Numerical values that determine the strength of connections between neurons in a neural network. These are adjusted during training to learn patterns from data.
Example: If edge patterns in cat images are useful, weights connected to those pixel combinations will be higher.
-
Short for Explainable AI, it refers to models or tools that offer human-understandable explanations for their decisions.
Example: An XAI system might explain why a loan was denied based on specific credit history indicators.
-
A fast object detection algorithm that processes the entire image in one go and identifies multiple objects in real-time.
Example: Used in self-driving cars to detect vehicles, pedestrians, and signs quickly and simultaneously.
-
A machine learning technique where the model can make predictions about classes it has never seen before, using semantic knowledge or descriptions.
Example: A model trained on dogs, cats, and birds might recognize a “hamster” based only on a textual description like “small rodent with cheek pouches.”
The easiest way to learn AI in 30 mins a day
Start learning AI for free by joining one of our free challenges.
💻 7 Days of AI Coding
Learn how to create MVPs and custom apps with AI in 7 days.
Join the challenge🤖 7 Days of AI Agents
Learn to build, customise and deploy AI Agents in just 7 days.
Join the challenge