AI Terms Glossary
116 quality-checked concepts spanning machine learning, foundation models, agents and AI safety—with clear English explanations.
A
Active Learning
Modelin is the learning strategy that the data samples that will be most useful to label.
Adversarial Attack
The method of adding small, intentional changes, as soon as a model will not be noticed with human eye.
Agentic AI
Not only answer the question of the model, but independently plan and complete multi-step tasks using the tool.
AI Agent
The presence of software that detects, decides and actions to achieve a specific goal.
AI Ethics
discipline of the development of artificial intelligence systems with fair, safe and social values.
AI Winter
The period sharply reduced due to the fact that interest and financing with artificial intelligence research is not met.
Application Programming Interface (API)
interface that allows different software to communicate with each other in a standard way.
Attention Mechanism
The mechanism that determines which parts of the input will focus more on when producing an output of the model.
Autoencoder
The nervous network that learns efficiently by compressing the input and learning to recreate again.
AutoML
Model selection, approach to automate processes such as hyperparameter adjustment.
Autonomous Vehicle
With computer vision, sensor fusion and decision algorithms, it can be diluted without human intervention.
B
Backpropagation
Training algorithm that updates weights by spreading the error in the nervous network from the output.
Tags
The size of the sample group given to the model at the same time during the training.
tool
The unjust trend reflected in model outputs of imbalances in education data.
Big Data
The concept that is large, fast-produced and expresses various data clusters until it cannot be processed by traditional methods.
C
Chain-of-Thought
The model’s ultimate response is the prompt technique that allows you to reach by step by step by step.
tool
The process of splitting a long document into small meaningful parts that can be processed.
Classification
The task of monitoring an entry into one of the predefined categories.
Closed Source Model
The proprietary model that does not share weights, accessible only via the API.
tool g
Unsupervised learning technique that separates similar data points into groups without tag.
Computer Vision
The area that allows computers to extract meaningful information from images and videos.
Confusion Matrix
Table showing correct and wrong predictions of a classification model according to classes.
Context Window
Maximum amount of tokens that a language model can work at once.
Convolutional Neural Network / CNN
The type of nervous network that learns by scanning the traces of raditional images with filters.
Cross-Validation
The method of measuring the overallization of the model by splitting the data set into different subsets.
D
Data Augmentation
The technique of replicating educational data by diversifying new variations from existing data.
Data Leakage
Due to the fact that the information of the test data is accidentally mixed in the educational process, the performance of the mod…
tool Tags
Collected, edited data samples collection for a specific task.
Decision Tree
Classification or predicting model structure by separating data into branches with a number of yes/nounce questions.
tool g
Using multi-layer artificial nerve networks, the machine learning that learns complex nucleus is lower.
Diffusion Model
Starting from random noise, step by step is a meaningful visual or sound producing manufacturer model type.
Digital Twin
Virtual copy updated with real-time data of a physical system.
Dimensionality Reduction
Reduce the number of variables in the data set without losing information as possible.
Close Tags
Regularization technique that reduces excessive learning by disabling random neurons during training.
E
Edge AI
The approach of the operation of the device itself (phone, camera, etc.) instead of the model's cloud.
tool g
The method that represents meaningal relations of words, phrases or objects as numerical vectors.
Emotion Diagnostic
technology to detect the emotional status of a person from facial expression, tone or text.
Coder Architectur News
The nerve network architecture, which first coded to a compressed representation, then generates output from this representation.
Learn More g
Methods that produce more powerful results than a single model by combining multiple models predictions.
Clearable AI / XAI
All of the methods that make decisions of a model understandable by the human beings.
F
Facial Recognition
Identify or verification technology by analyzing the facial features on a image or video.
tool
A measurable feature that forms the input of a model, identifies the data.
FeatureEngineer g
The process of deriving meaningful input variables from raw data to better learn the model.
Learn More g
The method of moving data to a central server, providing local model training on devices.
Learn More g
How to teach a new task by showing several examples to the model.
Fine-tuning
Retraining a pre-trained model with special data for a specific task.
Function Calling
Ability to suggest a language model to call a defined function with correct parameters.
Fuzzy Logic
The logic system that works with a degree of accuracy values instead of exact right/burning.
G
Generative Adversarial Network / GAN
The manufacturer of model architecture consisting of two different networks that compete against each other.
Gradient Tags
Optimisation algorithm step by step to minimize lost function.
Graphics Processing Unit / GPU
The type of processor widely used in parallel calculation, artificial intelligence training.
H
I
Image Segmentation
The task of a computerized view that distinguishes a image into meaningful areas at pixel level.
In-context Learnin g
The model learns only from examples given in prompt without updating weights.
Inference
The stage that a trained model produces prediction or output on new inputs.
J
K
K-Leck Neighbors / KNN
A data point is a simple algorithm that is classified according to most of the closest neighboring class.
Knowledge Distillation
The process of transferring the knowledge of a large and complex model to a smaller model.
L
Large Language Model / LLM
The artificial intelligence model, which is trained with Devasa text data, can produce text similar to human language.
Latent Space
The compressed of the data, the abstract multi-dimensional space representing the basic properties.
Long Short-Term Memory / LSTM
Advanced RNN type that can better learn dependencies in long series.
Loss Function
Function that measures the difference between model forecasts and actual values.
M
Machine Information g
All of the methods that allow systems to predict by learning the image from data without express programming.
Model Card
Standard document documenting capabilities, limitations and recommended usage areas of a model.
Model Collapse
The situation that models lose diversity and quality as a result of retraining with data made by artificial intelligence.
Model Compression
Integration of protection techniques as possible while reducing the size and calculation cost of a model.
Model Robustness
Ability to perform reliable performance in the face of a model, noisy, unexpected or intentionally manipulated inputs.
Multi-agent System
The system that multiple artificial intelligence agents work together in collaboration or competition.
Multimodal AI
Artificial intelligence system that can process different types of data such as text, visual, audio.
N
Naive Bayes
Based on Bayes teoremine, simple probability classifier, if attributes are independent of each other.
Named Entity Recognition / NER
NLP task that detects and classifies special names such as the person inside the text.
Natural Language Generation / NLG
The process of producing text similar to human language, fluent from structured data or a representation.
Natural Language Processing / NLP
The area that allows computers to understand, comment and produce human language.
Natural Language Understanding / NLU
The NLP sub-section, which allows a machine to grasp the meaning and intention of the human language.
Neural Network
The calculation model consisting of layer nodes, inspired by neuron connections in the human brain.
O
Online Learnin g
The model’s learning approach that is updated continuously and incrementally as data comes.
Open Source Model
The model of artificial intelligence that is clearly shared with weights and often code to everyone.
Optical Character Recognition / OCR
Technology that converts printed or handwriting text into editable digital text.
tool g
The situation where the model can not change the education data and do not generalize in new data.
P
Perplexity
Statistical metric that measures how well a language model predicts a text array.
tool / PCA
The statistical method that does not reduce the data to less size by finding dimensions that explain the maximum amount in the dat…
About us g
Art and science to design input texts (prompt) to get the desired output from artificial intelligence models.
R
Recall
metric measuring how much actual positive samples are located correctly by the model.
Recommendation System
The system recommends products or contents that may be interested in user behaviour.
Recurrent Neural Network / RNN
The type of nervous network that carries the knowledge of the previous steps to capture dependency in the sequence data.
Red Team g
Intently testing a AI system’s security open and unwanted behavior.
Regularization
The techniques that add additional constraints to the lost function to avoid excessive learning of the model.
Search g
The method that a agent learns optimal behavior through experimenting with the reward-ceza mechanism.
Reinforcement Learning from Human RLHF
The method of training that model is rewarded by human preferences.
Retrieval-Augmented Generation / RAG
The method that checks and binds the data about an external source of information when producing the model.
S
Self-Attention
The type of attention that calculates the relationship of each element in a number with all other items in the same directory.
Supervised Learnin g
Modelin learn from unlabeled data by predicting part of the data inside.
Semi-supervised Learnin g
Learning method that uses a small amount of labels and a large amount of unlabeled data together.
Speech Recognition
Technology that converts voice speech to written text.
Style Transfer
technique that implements the artistic style of another image while maintaining the contents of a image.
Superintelligence
In all areas of human intelligence, assumptional advanced artificial intelligence.
Supervised g
The method of learning the input-output relationship using the labeled data of the model.
Support Vector Machine / SVM
The surveillance learning algorithm that distinguishes data by finding the largest gap between classes.
Swarm Intelligence
collective intelligent behavior that occurs from the interaction of a large number of agents following simple rules.
T
tool News
The set that controls how random or predictable will behave when producing the model output.
Tensor Processing Unit / TPU
Developed by Google, specially designed accelerator chip for artificial intelligence workloads.
Text-to-3D Generation
Ability of artificial intelligence that produces three-dimensional models or scenes from a written description.
Text-to-Image Generation
The ability of artificial intelligence that produces a unique visual from a written description.
Text-to-Video Generation
The ability of artificial intelligence that produces movable, consistent video from a written description.
Token
The smallest meaning that a language model uses when processing the text.
Tool Use
The calculator to complete the task of a model can call external tools such as search engine.
Top-p / Top-kSample
sampling strategies that limit the potential distribution when choosing the next word.
Training Data
All examples used in the learning process of a model.
Transfer g
Transfer of information learned in a task into a different but associated task.
Transformer
Based on the attention mechanism, the structure of the nervous network, which constitutes the architectural basis of today’s large…
U
tool g
The situation that the model remains simple enough to learn the images within the data.
Unsupervised g
The method that the model explores hidden structure and images in unlabeled data.
V
Vector Database
The database type that allows you to store and search for meaningal similarity.
Vision-Language Model/VLM
Multi mode model type that can understand and associate both images.