Comprehensive AI Glossary: Key Terms and Concepts Explained

Understanding the Language of Artificial Intelligence for Beginners and Experts Alike

Hello, Alex from Questflow here.

The top buzzword of the world in 2022, 2023 and 2023 is currently in and around artificial intelligence. That's why we have created this AI Glossary. The AI space can be quite daunting and the terms are confusing, therefore we put together a glossary for both beginner and experienced folks to sharpen their verbal skills. We hope you find it useful. Of course, we'll add to it over time too as new terms are always a-coming. In that way, you would be able to catch up with the most recent AI ideas.

Terms:

Artificial Intelligence (AI): is known to simulate human intelligence in machines programmed to think like humans and mimic their actions. This includes autonomous machines that work on human cognition, such as learning, problem-solving, decision-making, and natural language processing.

Accelerator: A specialized piece of hardware designed to speed up the heavy computations required for tasks like deep learning and processing large datasets. Think of it as a supercharged engine that makes AI applications run much faster.

AI Agent: a computational entity that perceives its environment through sensors and acts upon that environment through effectors in order to achieve specific goals.

AGI (Artificial General Intelligence): is the hypothetical form of artificial intelligence that can understand or learn any intellectual task a human being does and further, able to solve problems in an unsupervised environment.

ASI (Artificial Super Intelligence): is an intelligence in all cognitive domains that will far surpass the best human intellects across every domain of cognition..

Backpropagation: It is a supervised learning algorithm used to adjust the weights of the network so as to minimize prediction errors in its output.

Bias: AI bias refers to built-in, automatic errors in AI models that can result in biased outcomes - results from pre-conceived or outright distorted beliefs within the data being processed by an A.I

Chain of Thought: is the capability to break complex asks into simpler inferences like humans do it.

ChatGPT: is an OpenAI model built to work on understanding and generating human-like text.

CLIP (Contrastive Language–Image Pretraining): is a model from OpenAI. It tries to understand and produce a link among visual data and text. CLIP uses an interesting training method.

Compute: The term "compute" refers to the computational resources and capabilities needed to train, deploy, and run AI models. Compute is crucial in AI. This is because processing large datasets and executing complex algorithms demand significant resources.

Convolutional Neural Network (CNN): A Convolutional Neural Network (CNN) is a deep learning model tailored for handling structured grid data, like images. 

Data Augmentation: Data Augmentation in AI means making more training data by changing existing data. This helps machine learning models, especially deep learning ones, work better. You can use it for different data types like images, text, and audio.

Deep Learning: is a subset of Artificial Intelligence and Machine Learning, called Deep learning. These are neural networks with many layers, so it is called “deep.” It helps the computer in understanding complicated data over many non-linear steps.

Diffusion: In AI, diffusion is a technique where we add random noise to data to create new data that's similar but not identical. It's often used in generative models to produce new samples that resemble the training data.

Double Descent: is a fascinating phenomenon observed during the model training processes, especially on models like deep learning. It explains why the error of a model goes down, then up, and again down as we double its complexity or after doubling substantial dataset has been added.

Embedding: refers to a technique for mapping data like words or images onto vectors in a continuous vector space where similar items are closer together.

Emergence/Emergent Behavior: Emergent behavior is characterized by complex systems and properties that emerge from the interactions of simpler entities within a system, exhibiting non-trivial behaviors or functions.

End-to-End Learning: This is a machine learning approach where the model learns directly from raw data to output without any intermediate steps. For example, it can go straight from audio recordings to text transcriptions without needing pre-processed features.

Expert Systems: These are AI programs designed to solve complex problems in a specific area, like medical diagnosis or financial forecasting, by mimicking the decision-making abilities of a human expert.

Explainable AI (XAI): XAI refers to the understandable decisions of Artificial Intelligence. It means whatever inside an XAI system should be human understandable, in which the humans can know why that particular decision was followed and how it reached there.

Fine-tuning: Adjusting a pre-trained language model on a smaller, specific dataset like customer service logs to improve performance on that task.

Forward Propagation: In artificial intelligence, and especially in neural networks, forward propagation is the process where input data propagate forward through the net's layers to generate an output. 

Foundation Model: A foundation model, as the term suggests, is a large neural network used to support various applications and tasks in AI. 

Generative Adversarial Network (GAN): is a class of machine learning framework. This involves two neural networks (the generator and the discriminator). The generator learns to create increasingly realistic data that can fool the discriminator, while the discriminator learns to better differentiate real data from fake.

Generative AI: is a subfield of artificial intelligence focused on creating content, text and information in particular, that tends to work like a human would. It uses advanced algorithms, including deep learning algorithms like variety, adversarial and generative networks and more to synthesize text, images, music and many other things. 

GPT (Generative Pretrained Transformer): an open-source neural network model introduced by OpenAI. The purpose of it is to produce text as humans, from the input. 

GPU (Graphics Processing Unit): GPUs perform parallel processing very well and this makes them especially great for training AI models and especially anything involving deep learning. 

Gradient Descent: is a common optimization algorithm used in machine learning and artificial intelligence to minimize the cost function which measures the difference between actual and predicted values. 

Hallucinate/Hallucination: is when a model produces outputs that have no basis in its training data or input. These outputs might be fake news, inaccuracies, or things that the AI is just as good as inventing themselves since they never are inputed. 

Hidden Layer: is a layer between input layers and output layers. The presence of hidden layers allows the capturing of complex patterns in the data, which results in more precise and powerful models.

Hyperparameter Tuning: is the process of finding the best hyperparameters for a machine learning model. Hyperparameter tuning in most cases leads to phenomenal improvements in the performance and accuracy of a model.

Inference: The process by which a trained machine learning model makes predictions or decisions based on new input data. It involves using mathematical algorithms and learned patterns from existing data. These algorithms help generate outcomes based on the new input.

Instruction Tuning: is the process of fine-tuning a model using particular instructions or examples that are intended to help teach the model how to perform better at given tasks. This meant feeding the AI an exercise in steps of increasing detail together with their expected responses. 

Large Language Model (LLM): is an advanced neural network that processes human language to learn, generate, and manipulate it. These trained models have seen tons of text data and thus can learn the patterns, context, and nuances of language. 

Latent Space: in AI, latent space is a type of multi-dimensional space that attempts to encode the input data in such a way that it allows for generalization or accurate prediction.

Loss Function (or Cost Function): A loss function is a feature that a machine learning model seeks to minimize during training. It quantifies how far the model's predictions are from the true values.

Machine Learning: is a part of Artificial Intelligence that deals with the development of algorithms and statistical models enabling computers to perform specific tasks without being explicitly programmed. Rather, they learn from the patterns and insights found within data.

Mixture of Experts: (MoE) in machine learning is a variety of machine learning model that builds on the idea that we can divide a problem into sub-tasks, solved by a specific expert model.

Multimodal: multimodal means systems or models that can process multiple types of data like text, image, sound and videos. A Multimodal AI uses extensive arrays of data types to interpret and convey comprehensions as well as enforce actions on different tasks.

Multi-agent framework: a system consisting of multiple intelligent agents that interact and collaborate with each other in the realization of complex tasks.

Natural Language Processing (NLP): is a subfield of artificial intelligence and machine learning concerned with the interactions between computers and human language.

NeRF (Neural Radiance Fields): A method in AI that uses a neural network to create detailed 3D scenes from multiple 2D images. It's used for things like photorealistic rendering and view synthesis.

Neural Network: is a computational model based on biological neural networks in the human brain. It is composed of one input layer of nodes, or neurons, followed by several layers interconnected with each other of nodes where each node represents a particular mathematical function.

Objective Function: is a mathematical formula used by the optimizer to evaluate and monitor the performance of an algorithm or model.

Overfitting: is the creation of too exact a model that is in support of the training data, but which will not perform well with new or future data. This occurs when the model is too complicated, having too many parameters with respect to the number of samples.

Parameters: term used in artificial intelligence (AI) to refer to the variables that define a model's configuration and behavior. These parameters are learned on data while training the model.

Pre-training: This is the initial phase where a model is trained on a large dataset to learn general features and patterns. After this, the model is fine-tuned on a smaller, more specific dataset to perform a particular task better.

Prompt: the input given to an AI model to prompt a response in artificial intelligence. This text can vary from a question, to a command, to a statement which is then used in training the AI to compose sensible and contextually sound sentences.

Regularization: In machine learning, regularization involves adding extra information to a model to prevent it from overfitting. This means it helps the model generalize better to new, unseen data by not relying too much on complex patterns in the training data.

Reinforcement Learning: is a brand of machine learning, where an agent learns to take action in an environment by receiving feedback from the environment. It learns through trial and errors so that it can maximize some cumulative reward.

RLHF (Reinforcement Learning from Human Feedback): is a subset of artificial intelligence (AI) that improves machine learning models by feeding them human feedback. With RLHF, humans provide feedback on when the AI does things or makes decisions in order to steer it towards better results.

Singularity: is a hypothetical point in the future related to the development of artificial intelligence (AI) at which technological growth becomes uncontrollable and irreversible and may result in unforeseeable changes to human civilization.

Supervised Learning: is another type of machine learning, completing a model with labeled output. This implies each training example has an output label associated. It is trained with many examples of input-output pairs so that it learns to map inputs to the correct output.

Symbolic Artificial Intelligence: is an AI approach which takes the logic as its basis and uses symbols for representing problems and knowledge.

TensorFlow: is the open-source machine learning framework that was developed by Google. It offers an end-to-end ecosystem for building, training, and deploying machine learning and deep learning models.

TPU (Tensor Processing Unit): is an application-specific integrated circuit developed by Google specifically for accelerating machine learning workloads, especially neural networks.

Training Data: This is the dataset that teaches a machine learning model. It includes pairs of data: the input data that goes into the model and the output data that the model is supposed to learn to predict.

Transfer Learning: is a technique used in artificial intelligence (AI) where a model designed for one task is re-purposed on a second related task.

Transformer: is a deep learning model architecture that use self-attention mechanisms to process inputs, letting the model weigh input sequences differently in context. This architecture is capable of learning long-range dependencies more efficiently than a traditional RNN.

Underfitting: is when a model is too basic to capture the data that was used to train it. This frequently happens when the model is not complex enough, (in the case of Neural networks), this means small number or layers or a few numbers of parameters.

Unsupervised Learning: A type of machine learning where the model is not provided with labeled training data. Instead, the model must identify patterns in the data on its own.

Validation Data: A distinct subset of the dataset utilized in machine learning, which is distinct from both the training and test datasets. It is utilized to optimize the hyperparameters

XAI (Explainable AI) is a subfield of artificial intelligence dedicated to creating transparent models. These models provide clear and understandable explanations of their decisions.

Zero-shot Learning: is a type of inductive learning that aims to recognize and categorize objects or data-points for which it has not received training instances.

Some words related to Agents:

Action: A specific task performed by an AI agent within its environment. Actions are how the AI executes tasks and interacts with its surroundings.

Agent: “agent” in the context of artificial intelligence refers to an autonomous semantic entity perceiving its environment with the help of sensors and acting through actuators.

Agent Function: Agent Function: A mathematical expression that maps an agent's percept history (input information about the current state of the world) to an output action.

Agent Program: is a software object that perceives its environment through sensors and acts upon it using actuators. These are intelligent, self-dependent software systems that can function without any human involvement; they take decisions on the basis of rules or strategies defined already, derived only from previous experiences.

Environment:Everything that surrounds and interacts with an AI agent. It includes all external factors that influence the agent's behavior and decision-making.

Goal: symbolizes a result which AI system aims to achieve through the operations and decision making. 

Learning Agent: is a system that learns to perform better with time from its experiences. Learning agents are agents that learn from data generated by the system instead of directly following significant predefined rules.

Multi-Agent Systems (MAS): are systems containing multiple interacting intelligent agents working in a shared environment to solve individual or collective problems.

Perception: refers to the result of a process by AI systems that interprets sensory input coming from its environment. This includes the input of data from cameras (visual perception), microphones (audio perception), and other sensors.

Policy: A set of rules or strategies an agent follows when taking actions to achieve its goal

Rational Agent: is anything which acts on the basis of past input and aims to maximize its behavior. 

State: A specific condition or situation an AI system or agent can be in at any moment. It includes all the information needed to describe the system's status within its environment.

Utility: A measure of satisfaction or preference for an outcome. It is a key concept in decision theory and economics that guides AI decision-making processes.

Connect with Questflow: Website | X | Telegram | Discord | GitHub