AI and Machine Learning Glossary for IT-Officer Exams

176 terms in plain English, one or two sentences each, with the formulas exams ask for. Search, jump by letter, or filter by topic. Terms that fall under a topic named in the PFRDA or UGC NET Computer Science syllabus carry that exam's tag.

  • 176 terms
  • 11 topics
  • 92 tagged PFRDA
  • 66 tagged UGC NET

Why it matters

Where AI turns up in IT-officer exams

Each row uses the syllabus wording recorded on our page for that exam.

AI in IT-officer exam syllabuses
ExamWhat the syllabus saysMore
PFRDA Officer Grade A 2026General-stream Paper 2, Section J: Artificial Intelligence / Machine Learning. Named topics include supervised and unsupervised learning, data preprocessing, model evaluation, NLP, sentiment analysis, overfitting, LLMs, agentic AI, responsible AI, cross-validation and reinforcement learning.PFRDA page
UGC NET Computer SciencePaper II, Unit 10: Artificial Intelligence. Approaches to AI, knowledge representation, planning, NLP, multi-agent systems, fuzzy sets, genetic algorithms and artificial neural networks.UGC NET CS page
IB DCIO / TechArtificial intelligence is one of the thirteen syllabus topics. UPSC gives no sub-topics or marks for it.DCIO page
UPSC Assistant Director (Systems)Artificial Intelligence and Machine Learning is a syllabus topic.AD (Systems) page
IBPS SO IT OfficerIBPS publishes no detailed syllabus. Our guide lists AI and machine learning basics under emerging technology.IBPS SO IT page
IBPS RRB SO ITPast-paper topic lists include artificial intelligence and data mining questions.RRB SO IT PYQs
Syllabus:

176 terms

A

Accuracy
The share of all predictions that are correct: (TP + TN) ÷ total. It misleads when one class is rare, for example in fraud detection.
Model evaluationPFRDA
Activation Function
The non-linear function applied to a neuron's weighted sum. Sigmoid gives outputs in (0, 1), tanh in (−1, 1), and ReLU gives max(0, x).
Neural networks and deep learningUGC NET
Admissible Heuristic
A heuristic that never overestimates the true cost to reach the goal. Straight-line distance on a road map is the textbook example.
Search and game playingUGC NET
Agentic AI(AI agents)
AI systems, usually built on an LLM, that pursue a goal over several steps on their own: planning, calling tools such as search or code, checking results and deciding what to do next.
Generative AI and LLMsPFRDA
Algorithmic Bias
When an AI system produces unfair results for some groups, often because its training data under-represents them or reflects past discrimination.
Responsible AIPFRDA
Alpha-Beta Pruning
An improvement to minimax that skips branches which cannot change the final decision. It returns the same move as minimax, and with good move ordering can search about twice as deep in the same time.
Search and game playingUGC NET
Artificial General Intelligence(AGI, Strong AI)
A hypothetical AI that could learn and perform any intellectual task a human can, across domains. No such system exists yet.
AI basics
Artificial Intelligence(AI)
The field of building computer systems that do tasks which normally need human intelligence, such as reasoning, learning, understanding language and recognising images.
AI basicsUGC NET
Artificial Neural Network(ANN)
A model made of layers of connected units (neurons), each computing a weighted sum of its inputs and passing it through an activation function. The weights are learned from data.
Neural networks and deep learningUGC NET
Association Rule
A rule of the form X → Y found in transaction data, such as "bread → butter". Support is the share of transactions containing both X and Y; confidence is support(X ∪ Y) ÷ support(X). Apriori is the standard algorithm.
AI basics
Attention Mechanism
A way for a model to weigh how relevant every other token is when processing each token. In self-attention, each token is compared with all tokens in the same sequence.
Generative AI and LLMsPFRDA
Autoencoder
A network trained to reproduce its input through a narrow middle layer, so it learns a compressed representation. Used for dimensionality reduction and anomaly detection.
Neural networks and deep learning

B

Backpropagation
The algorithm that computes the gradient of the loss with respect to every weight by applying the chain rule backwards from the output layer, so gradient descent can update the weights.
Neural networks and deep learningUGC NET
Backward Chaining
Goal-driven reasoning that starts from the goal and works backwards, looking for rules and facts that would prove it. Prolog works this way.
Knowledge, logic and planningUGC NET
Bag of Words
Representing a document by how often each word appears, ignoring grammar and word order.
Natural language processingPFRDA
Bagging
Bootstrap aggregating: train models in parallel on random samples drawn with replacement, then average or vote. It mainly reduces variance.
Machine learningPFRDA
BERT
Bidirectional Encoder Representations from Transformers: Google's 2018 encoder-only model, pre-trained by predicting hidden (masked) words using context on both sides. Suited to understanding tasks such as classification.
Generative AI and LLMsPFRDA
Bias-Variance Trade-off
High bias (too simple) causes underfitting; high variance (too sensitive to the training data) causes overfitting. Reducing one often raises the other.
Model evaluationPFRDA
Big Data
Data too large or fast-changing for traditional tools, usually described by the 3 Vs: volume, velocity and variety (often extended with veracity and value).
AI basics
Boosting
Train models one after another, each focusing on the errors of the previous ones, then combine them. AdaBoost, gradient boosting and XGBoost are examples. It mainly reduces bias.
Machine learningPFRDA

C

Chain-of-Thought Prompting
Asking a model to work through intermediate reasoning steps before giving the answer, which often improves accuracy on multi-step problems.
Generative AI and LLMsPFRDA
Chatbot
A program that holds a conversation with users in text or speech. Early chatbots such as ELIZA (1966) matched patterns; modern ones are built on large language models.
AI basics
Chromosome Encoding
How a candidate solution is written for a genetic algorithm, for example as a binary string, a permutation or a list of real values.
Fuzzy logic and genetic algorithmsUGC NET
Classification
Supervised learning that predicts a category, such as fraud or not fraud.
Machine learningPFRDA
Clustering
Unsupervised learning that groups similar examples together without labels.
Machine learningPFRDA
Computer Vision
The field of AI that lets computers interpret images and video, for example to detect objects, read text or recognise faces.
AI basics
Conceptual Dependency
Roger Schank's way of representing the meaning of sentences with a small set of primitive acts, such as PTRANS (physical transfer), so that sentences with the same meaning get the same representation.
Knowledge, logic and planningUGC NET
Confusion Matrix
A table of a classifier's results: true positives (TP), false positives (FP), false negatives (FN) and true negatives (TN).
Model evaluationPFRDA
Constraint Satisfaction Problem(CSP)
A problem defined by variables, the values each may take (domains) and constraints between them, such as map colouring or Sudoku. It is usually solved by backtracking with constraint propagation.
Search and game playing
Context Window
The maximum number of tokens a model can take into account at once, covering both the input and its own output.
Generative AI and LLMsPFRDA
Convolutional Neural Network(CNN)
A network that slides small learned filters over an image to detect features such as edges and shapes, usually followed by pooling layers. The standard model for images.
Neural networks and deep learning
Cross-Validation
Estimating performance by splitting the data into k parts (folds), training on k − 1 and testing on the remaining one, k times, then averaging the scores. Leave-one-out uses k equal to the number of examples.
Model evaluationPFRDA
Crossover
A genetic operator that combines parts of two parent solutions to make offspring, for example by swapping the bits after a random cut point.
Fuzzy logic and genetic algorithmsUGC NET

D

Data Leakage
When information from outside the training data, often from the test set or the future, slips into training. It makes a model look better in testing than it will be in real use.
Machine learningPFRDA
Data Mining
Finding useful patterns, such as associations, clusters or anomalies, in large data sets. It is one step of knowledge discovery in databases (KDD).
AI basics
Data Preprocessing
Preparing raw data for a model: cleaning errors, handling missing values, encoding categories, scaling numbers and splitting the data.
Machine learningPFRDA
Data Privacy in AI
Protecting personal data used to train or run AI systems, through consent, minimising data, anonymisation and secure storage. In India, the Digital Personal Data Protection Act, 2023 applies.
Responsible AIPFRDA
Decision Tree
A model that splits data by asking one feature question at each node until it reaches a leaf with a prediction. Splits are chosen by information gain (entropy) or Gini impurity.
Machine learningPFRDA
Deep Learning
Machine learning with neural networks of many layers, which learn useful features directly from raw data such as pixels, audio or text.
AI basicsPFRDA
Deepfake
Realistic fake video, images or audio of a real person, made with generative AI. A growing tool for fraud and misinformation.
Responsible AIPFRDA
Defuzzification
Converting the fuzzy output of a rule system back into one crisp value. The centroid (centre of gravity) method is the most common.
Fuzzy logic and genetic algorithmsUGC NET
Diffusion Model
A generative model that learns to reverse a process of gradually adding noise, so it can turn random noise into a new image step by step.
Generative AI and LLMs
Dimensionality Reduction
Reducing the number of features while keeping most of the useful information, to speed up learning, reduce noise or allow plotting.
Machine learningPFRDA
Dropout
A regularisation method that switches off a random share of neurons at each training step, so the network cannot rely on any single one.
Neural networks and deep learning

E

Embedding
A vector of numbers that represents the meaning of a piece of text, an image or another item, so that similar items have nearby vectors.
Generative AI and LLMsPFRDA
Ensemble Learning
Combining several models to get better predictions than any one alone. Bagging, boosting and stacking are the main approaches.
Machine learningPFRDA
Entropy and Information Gain
Entropy measures impurity: H = −Σ p log₂ p, which is 0 for a pure node and 1 for a 50:50 split of two classes. Information gain is the drop in entropy after a split; ID3 picks the split with the highest gain.
Machine learningPFRDA
Epoch and Batch
An epoch is one full pass over the training data. A batch is the subset of examples used for one weight update; with 1,000 examples and batches of 100, one epoch has 10 updates.
Neural networks and deep learning
Expert System
A program that solves problems in a narrow field using a knowledge base of facts and if-then rules plus an inference engine that applies them. MYCIN, for diagnosing blood infections, is the classic example.
AI basicsUGC NET
Explainable AI(XAI)
Methods that make a model's decisions understandable to people, such as showing which features drove a prediction. LIME and SHAP are common tools.
Responsible AIPFRDA
Exploration vs Exploitation
The trade-off between trying new actions to learn more (exploration) and using the best action known so far (exploitation). ε-greedy explores at random with a small probability ε.
Reinforcement learningPFRDA

F

F1 Score
The harmonic mean of precision and recall: 2 × P × R ÷ (P + R). Useful when classes are imbalanced.
Model evaluationPFRDA
Feature
One measurable input property used by a model, such as age, income or a pixel value. A row of features describes one example.
Machine learningPFRDA
Feature Engineering
Creating or transforming features to help a model learn, such as extracting the month from a date.
Machine learningPFRDA
Fine-tuning
Further training of a pre-trained model on a smaller, task-specific data set to adapt it to that task or style.
Generative AI and LLMsPFRDA
First-Order Logic(Predicate logic)
Logic with objects, predicates, functions and the quantifiers ∀ (for all) and ∃ (there exists), so it can state facts like "every officer passed an exam".
Knowledge, logic and planningUGC NET
Fitness Function
The function that scores how good a candidate solution is in a genetic algorithm. Fitter candidates are more likely to be selected as parents.
Fuzzy logic and genetic algorithmsUGC NET
Forward Chaining
Data-driven reasoning that starts from known facts and applies rules to derive new facts until the goal is reached or nothing new follows.
Knowledge, logic and planningUGC NET
Foundation Model
A large model pre-trained on broad data that can be adapted to many different tasks. LLMs are the best-known example.
Generative AI and LLMs
Frame
A record-like structure for a stereotyped object or situation, made of slots and their fillers, with default values and inheritance. Proposed by Marvin Minsky.
Knowledge, logic and planningUGC NET
Fuzzification
Converting a crisp input value into degrees of membership in fuzzy sets, the first step of a fuzzy inference system.
Fuzzy logic and genetic algorithmsUGC NET
Fuzzy Inference System
A system that fuzzifies inputs, applies IF-THEN fuzzy rules and defuzzifies the result. Mamdani and Sugeno are the two main types.
Fuzzy logic and genetic algorithmsUGC NET
Fuzzy Set
A set in which each element belongs to a degree between 0 and 1, instead of fully in or fully out. Introduced by Lotfi Zadeh in 1965.
Fuzzy logic and genetic algorithmsUGC NET
Fuzzy Set Operations
The standard fuzzy operations are union = max(μA, μB), intersection = min(μA, μB) and complement = 1 − μA.
Fuzzy logic and genetic algorithmsUGC NET

G

Generative Adversarial Network(GAN)
Two networks trained against each other: a generator makes fake samples and a discriminator tries to tell them from real ones, until the fakes become convincing. Introduced by Ian Goodfellow in 2014.
Generative AI and LLMs
Generative AI
AI that creates new content, such as text, images, audio or code, in response to a prompt, rather than only classifying or predicting.
Generative AI and LLMs
Genetic Algorithm(GA)
A search method inspired by natural selection: a population of encoded candidate solutions is improved over generations through selection, crossover and mutation.
Fuzzy logic and genetic algorithmsUGC NET
Goal Stack Planning
A linear planning method that keeps goals and actions on a stack and solves the goal on top first. It can fail on interacting goals, as in the Sussman anomaly.
Knowledge, logic and planningUGC NET
GPT
Generative Pre-trained Transformer: OpenAI's family of decoder-only transformer models, pre-trained to predict the next token and then tuned to follow instructions.
Generative AI and LLMsPFRDA
Gradient Descent
An optimisation method that repeatedly moves the weights a small step against the gradient of the loss: w ← w − η∇L, where η is the learning rate. Stochastic and mini-batch versions use part of the data per step.
Neural networks and deep learningUGC NET

H

Hallucination
When a generative model states something false or made up as if it were fact, such as a non-existent source.
Generative AI and LLMsPFRDA
Heuristic
A rule of thumb that estimates how close a state is to the goal, written h(n). A good heuristic lets a search expand far fewer nodes.
Search and game playingUGC NET
Hierarchical Clustering
Clustering that builds a tree of clusters (a dendrogram), either by merging the closest clusters step by step (agglomerative) or by splitting from one big cluster (divisive).
Machine learningPFRDA
Hill Climbing
Local search that keeps moving to the best neighbouring state. It is fast but can get stuck on a local maximum, a plateau or a ridge.
Search and game playingUGC NET
Hopfield Network
A recurrent network with symmetric weights that works as an associative memory: given a noisy or partial pattern, it settles into the closest stored pattern by lowering an energy function.
Neural networks and deep learningUGC NET
Human-in-the-Loop
A design in which people review, correct or approve an AI system's outputs, especially for high-stakes decisions such as loans.
Responsible AIPFRDA
Hyperparameter
A setting chosen before training, not learned from data, such as the learning rate, the k in k-NN or the depth of a tree. Tuned with a validation set or cross-validation.
Model evaluationPFRDA

I

Inference Engine
The part of an expert system that applies rules to the knowledge base to draw conclusions, using forward or backward chaining.
Knowledge, logic and planningUGC NET
Intelligent Agent
Anything that perceives its environment through sensors and acts on it through actuators. Exam questions often describe an agent by its PEAS: Performance measure, Environment, Actuators and Sensors.
AI basicsUGC NET

K

k-Means Clustering
Clustering that picks k centres, assigns each point to its nearest centre, moves each centre to the mean of its points, and repeats until nothing changes. k must be chosen in advance, often with the elbow method.
Machine learningPFRDA
k-Nearest Neighbours(k-NN)
A classifier that labels a new point by the majority class of its k closest training points. It is a lazy learner: there is no training step, and all work happens at prediction time.
Machine learningPFRDA
Knowledge Representation
How facts about the world are stored so a program can reason with them. Common forms are logic, semantic networks, frames, rules and scripts.
Knowledge, logic and planningUGC NET

L

Label
The correct output attached to a training example in supervised learning, such as "spam" or a house price.
Machine learningPFRDA
Large Language Model(LLM)
A very large neural network, usually a transformer, trained on huge amounts of text to predict the next token. It can then answer questions, summarise, translate and write code.
Generative AI and LLMsPFRDA
Learning Rate
The step size in gradient descent. Too large and training overshoots or diverges; too small and it is very slow.
Neural networks and deep learningUGC NET
Lemmatisation
Reducing a word to its dictionary form using vocabulary and grammar, so "studies" becomes "study" and "better" becomes "good". Slower but more accurate than stemming.
Natural language processingPFRDA
Linear Regression
A regression model that fits a straight line (or plane), y = w·x + b, usually by minimising the sum of squared errors (least squares).
Machine learningPFRDA
Linguistic Variable
A variable whose values are words, such as temperature = low, medium or high, each defined by a fuzzy set.
Fuzzy logic and genetic algorithmsUGC NET
Logistic Regression
A classification model that passes a linear score through the sigmoid function to give a probability between 0 and 1. Despite its name, it is used for classification.
Machine learningPFRDA
Long Short-Term Memory(LSTM)
A recurrent cell with input, forget and output gates that control what is kept, so it can learn long-range dependencies that a plain RNN loses.
Neural networks and deep learning
Loss Function
The measure of how wrong a model's predictions are, which training tries to minimise. Mean squared error for regression; cross-entropy for classification.
Neural networks and deep learningPFRDAUGC NET

M

Machine Learning(ML)
The part of AI in which systems learn patterns from data instead of following hand-written rules. Tom Mitchell's definition: a program learns from experience E on task T if its performance P on T improves with E.
AI basicsPFRDA
Markov Decision Process(MDP)
The standard model for reinforcement learning: states, actions, transition probabilities, rewards and a discount factor. The Markov property means the next state depends only on the current state and action.
Reinforcement learningPFRDA
Mean Squared Error(MSE)
The average of the squared differences between predicted and actual values. Squaring punishes large errors more. Its square root is RMSE.
Model evaluationPFRDA
Membership Function
The function μ(x) that gives each element's degree of membership in a fuzzy set, from 0 to 1. Triangular and trapezoidal shapes are common.
Fuzzy logic and genetic algorithmsUGC NET
Minimax
The decision rule for two-player, zero-sum games: the maximising player picks the move with the highest value, assuming the opponent will always reply with the move that minimises it.
Search and game playingUGC NET
Missing Value Imputation
Filling in missing data, for example with the column's mean, median or most frequent value, or with a value predicted from other columns.
Machine learningPFRDA
ML Model
The learned function that maps inputs to outputs, defined by its algorithm and the parameter values found during training. Building one usually means: collect data, prepare it, train, evaluate, then deploy.
Machine learningPFRDA
Model Drift
The decline in a deployed model's accuracy as real-world data changes from the data it was trained on. Models need monitoring and retraining.
Responsible AI
Model Evaluation
Measuring how well a trained model performs on data it has not seen, with metrics suited to the task: accuracy, precision, recall and F1 for classification; MSE, MAE and R² for regression.
Model evaluationPFRDA
Multi-Agent System
A system of several agents that act in the same environment and may cooperate, compete or negotiate, usually by passing messages in an agreed communication language.
AI basicsUGC NET
Multilayer Perceptron(MLP)
A feed-forward network with one or more hidden layers between input and output. With non-linear activations it can learn problems like XOR that a single perceptron cannot.
Neural networks and deep learningUGC NET
Multimodal AI
AI that works with more than one kind of data, such as text and images together, for example answering questions about a photo.
Generative AI and LLMs
Mutation
A genetic operator that makes a small random change to a solution, such as flipping one bit, to keep diversity in the population.
Fuzzy logic and genetic algorithmsUGC NET

N

N-gram
A sequence of n consecutive words or characters. Bigrams (n = 2) and trigrams (n = 3) capture a little word order.
Natural language processingPFRDA
Naive Bayes
A classifier based on Bayes' theorem that assumes features are independent of each other given the class. Simple and fast, and long used for spam filtering.
Machine learningPFRDA
Named Entity Recognition(NER)
Finding and labelling names of people, organisations, places, dates and amounts in text.
Natural language processingPFRDA
Narrow AI(Weak AI)
AI built for one task or a small set of tasks, such as spam filtering or face recognition. Every AI system in use today is narrow AI.
AI basics
Natural Language Processing(NLP)
The field of AI that lets computers read, understand and generate human language, from spell-checking and translation to chatbots.
Natural language processingPFRDAUGC NET
Normalisation
Rescaling a feature to a fixed range, usually 0 to 1, with x′ = (x − min) ÷ (max − min). Called min-max scaling.
Machine learningPFRDA

O

One-Hot Encoding
Turning a categorical feature into separate 0/1 columns, one per category, so a model does not read a false order into category codes.
Machine learningPFRDA
Ontology
A formal, shared description of the concepts in a domain and the relations between them. The semantic web describes ontologies in languages such as OWL.
Knowledge, logic and planningUGC NET
Overfitting
When a model learns the training data too closely, including its noise, so it scores well in training but badly on new data. More data, regularisation or a simpler model help.
Model evaluationPFRDA

P

Parsing
Analysing a sentence's grammatical structure against a grammar, usually producing a parse tree. Parsers work top-down or bottom-up.
Natural language processingUGC NET
Part-of-Speech Tagging(POS tagging)
Labelling each word in a sentence with its grammatical role, such as noun, verb or adjective.
Natural language processingUGC NET
Partial-Order Planning
Planning that fixes the order of two actions only when it must, so independent steps stay unordered. It handles interacting goals better than linear planning.
Knowledge, logic and planningUGC NET
Perceptron
The simplest neural unit: it outputs 1 if the weighted sum of its inputs passes a threshold, else 0. A single-layer perceptron can only learn linearly separable functions, so it cannot learn XOR.
Neural networks and deep learningUGC NET
Planning
Finding a sequence of actions that takes the world from an initial state to a goal state. Linear planners solve one goal at a time; non-linear planners interleave goals.
Knowledge, logic and planningUGC NET
Policy
The agent's rule for choosing an action in each state. Learning a good policy is the aim of reinforcement learning.
Reinforcement learningPFRDAUGC NET
Pragmatics
The study of meaning in context: what a speaker intends, as with "Can you pass the salt?", which is a request, not a question about ability.
Natural language processingUGC NET
Pre-training
The first, large-scale training of a model on general data, such as predicting the next token across a huge text collection, before it is adapted to specific tasks.
Generative AI and LLMsPFRDA
Precision
Of the cases the model called positive, the share that really are: TP ÷ (TP + FP).
Model evaluationPFRDA
Principal Component Analysis(PCA)
Dimensionality reduction that finds new, uncorrelated axes (principal components) ordered by how much of the data's variance they capture.
Machine learningPFRDA
Production Rule
A knowledge unit of the form IF condition THEN action. A set of such rules with a working memory and a control strategy forms a production system.
Knowledge, logic and planningUGC NET
Prompt Engineering
Designing the input given to a generative model, with instructions, context and examples, to get more accurate and useful output.
Generative AI and LLMsPFRDA
Prompt Injection
An attack that hides instructions in content a model reads, such as a web page or email, to make it ignore its original instructions or leak data.
Generative AI and LLMsPFRDA
Propositional Logic
Logic over whole statements that are true or false, joined by connectives such as AND (∧), OR (∨), NOT (¬) and implies (→). It cannot talk about objects or their properties inside a statement.
Knowledge, logic and planningUGC NET

Q

Q-Learning
A model-free reinforcement learning method that learns the value Q(s, a) of taking action a in state s, updating it from each reward it receives. It is off-policy.
Reinforcement learningPFRDA

R

R-squared(R², coefficient of determination)
The share of the variance in the target that a regression model explains, usually between 0 and 1. Higher is better.
Model evaluationPFRDA
Random Forest
An ensemble of decision trees, each trained on a bootstrap sample of the data and a random subset of features. It predicts by majority vote or averaging.
Machine learningPFRDA
Rational Agent
An agent that, for every situation, chooses the action expected to maximise its performance measure given what it has perceived so far. Rational means doing the best expected thing, not being all-knowing.
AI basicsUGC NET
Recall
Of the real positive cases, the share the model found: TP ÷ (TP + FN). Also called sensitivity or the true positive rate.
Model evaluationPFRDA
Recurrent Neural Network(RNN)
A network for sequences that passes a hidden state from one step to the next, so earlier inputs influence later outputs. Used for text, speech and time series.
Neural networks and deep learning
Regression
Supervised learning that predicts a number, such as a loan amount or a price.
Machine learningPFRDA
Regularisation
Adding a penalty for large weights to the loss to reduce overfitting. L1 (lasso) can shrink weights to exactly zero; L2 (ridge) shrinks them all smoothly.
Model evaluationPFRDA
Reinforcement Learning(RL)
Learning by trial and error: an agent takes actions in an environment and learns, from rewards and penalties, a policy that maximises total reward over time.
Reinforcement learningPFRDAUGC NET
Resolution
An inference rule used for proof by contradiction: convert the facts and the negated goal to clause form (CNF), then resolve clauses until the empty clause appears, which proves the goal.
Knowledge, logic and planningUGC NET
Responsible AI
Building and using AI so that it is fair, transparent, accountable, safe and respects privacy, with human oversight of important decisions.
Responsible AIPFRDA
Retrieval-Augmented Generation(RAG)
A method in which relevant documents are first retrieved, often from a vector database, and added to the prompt, so the model answers from those sources instead of memory alone.
Generative AI and LLMsPFRDA
Reward
The number the environment gives the agent after an action, telling it how good the result was. The agent tries to maximise the total, often discounted, reward.
Reinforcement learningPFRDAUGC NET
ROC Curve and AUC
The ROC curve plots the true positive rate against the false positive rate at every threshold. The area under it (AUC) is 0.5 for random guessing and 1.0 for a perfect classifier.
Model evaluationPFRDA

S

Script
A frame-like structure for a stereotyped sequence of events, such as going to a restaurant, with roles, props, scenes and results. Associated with Roger Schank.
Knowledge, logic and planningUGC NET
Self-Organising Map(SOM, Kohonen map)
An unsupervised network, introduced by Teuvo Kohonen, that maps high-dimensional data onto a low-dimensional grid by competitive learning, keeping similar inputs close together.
Neural networks and deep learningUGC NET
Self-Supervised Learning
Learning in which the labels are created from the data itself, for example by hiding a word and asking the model to predict it. Large language models are pre-trained this way.
Machine learning
Semantic Analysis
Working out the meaning of words and sentences, beyond their grammar, for example resolving which sense of "bank" is meant.
Natural language processingUGC NET
Semantic Network
A graph that represents knowledge as nodes for concepts and labelled edges for relations such as is-a and has-a. Properties can be inherited along is-a links.
Knowledge, logic and planningUGC NET
Semi-Supervised Learning
Learning from a small amount of labelled data together with a large amount of unlabelled data.
Machine learning
Sentiment Analysis(Opinion mining)
Classifying the opinion in text as positive, negative or neutral, for example in customer reviews or social media posts.
Natural language processingPFRDA
Simulated Annealing
Local search that sometimes accepts a worse move, with a probability that falls over time, so it can escape local maxima that trap hill climbing.
Search and game playing
Softmax
A function that turns a vector of scores into probabilities that are all positive and add up to 1. Used in the output layer for multi-class classification.
Neural networks and deep learningUGC NET
Specificity
Of the real negative cases, the share correctly called negative: TN ÷ (TN + FP).
Model evaluationPFRDA
Standardisation
Rescaling a feature to mean 0 and standard deviation 1, with z = (x − mean) ÷ standard deviation. Called z-score scaling.
Machine learningPFRDA
State Space
The set of all states a problem can be in, with the actions that move between them. Solving the problem means searching this space for a path from the start state to a goal state.
Search and game playingUGC NET
Stemming
Cutting words down to a crude root by stripping endings, so "studies" and "studying" both become "studi". Fast but can produce non-words. Porter's algorithm is the best known.
Natural language processingPFRDA
Stop Words
Very common words such as "the", "is" and "of" that carry little meaning and are often removed before analysis.
Natural language processingPFRDA
STRIPS
A classic planning language in which each action has preconditions, an add list and a delete list. Named after the Stanford Research Institute Problem Solver.
Knowledge, logic and planningUGC NET
Supervised Learning
Learning from labelled examples, where each input comes with the correct output, so the model can predict outputs for new inputs. Classification and regression are the two main kinds.
Machine learningPFRDAUGC NET
Support Vector Machine(SVM)
A classifier that finds the boundary with the widest margin between classes. The kernel trick lets it draw curved boundaries by working in a higher-dimensional space.
Machine learningPFRDA

T

Temperature
A setting that controls how random a model's output is. Low temperature gives focused, repeatable answers; high temperature gives more varied ones.
Generative AI and LLMsPFRDA
TF-IDF
A word weight that is high when a word is frequent in one document but rare across all documents: tf × log(N ÷ df), where N is the number of documents and df the number containing the word.
Natural language processingPFRDA
Token
The unit of text an LLM reads and writes, often a word or part of a word. Model limits and prices are counted in tokens.
Generative AI and LLMsPFRDA
Tokenisation
Splitting text into units called tokens, such as words or pieces of words, before processing.
Natural language processingPFRDA
Training, Validation and Test Sets
The usual split of data: the training set fits the model, the validation set tunes choices such as hyperparameters, and the test set gives a final, unbiased score.
Machine learningPFRDA
Transfer Learning
Reusing a model trained on one task as the starting point for a related task, which needs far less data and training time.
Machine learning
Transformer
The neural network design behind modern LLMs, introduced in the 2017 paper "Attention Is All You Need". It uses self-attention instead of recurrence, so it can process a whole sequence in parallel.
Generative AI and LLMsPFRDA
Turing Test
A test proposed by Alan Turing in 1950: if a human judge chatting in text cannot reliably tell a machine from a person, the machine is said to show intelligent behaviour.
AI basicsUGC NET

U

Underfitting
When a model is too simple to capture the pattern in the data, so it scores badly on both training and new data.
Model evaluationPFRDA
Unsupervised Learning
Learning from data without labels, to find structure such as clusters, associations or a lower-dimensional summary.
Machine learningPFRDAUGC NET

V

Vanishing Gradient
When gradients shrink as they pass back through many layers, so early layers barely learn. Sigmoid and tanh activations make it worse; ReLU and LSTM cells help.
Neural networks and deep learning
Vector Database
A database that stores embeddings and finds the most similar ones to a query quickly. Used to look up relevant passages for RAG.
Generative AI and LLMsPFRDA

W

Word Embedding
A dense vector of numbers representing a word, learned so that words used in similar contexts get similar vectors. Word2Vec and GloVe are well-known methods.
Natural language processingPFRDA

Z

Zero-Shot and Few-Shot Learning
Zero-shot: the model performs a task from an instruction alone, with no examples. Few-shot: a handful of worked examples are included in the prompt.
Generative AI and LLMsPFRDA

Definitions are written by The Victory Key for exam revision. Exam tags show which syllabus topic a term falls under; they do not mean the term has been asked in a past paper. Spotted an error? Tell us through the contact page.