Glossary of Artificial Intelligence

A

Artificial Intelligence (AI)

Artificial intelligence is the broad field of computer science dedicated to creating systems capable of tasks that typically require human intelligence. In practice, AI systems mimic cognitive functions such as learning from data, recognizing patterns, understanding language, and making decisions. Modern AI encompasses techniques like machine learning and deep learning to build smart software and robots that can adapt and improve over time.

Artificial General Intelligence (AGI)

This term refers to a hypothetical future AI that possesses human-level cognitive abilities across a wide range of tasks. Unlike today's narrow AI systems, which excel only at specific tasks, an AGI could autonomously understand, learn, and apply knowledge to any problem much like a human being. The concept of AGI contrasts with Artificial Narrow Intelligence (ANI) (also known as weak AI), which is limited to particular domains or functions, and with Artificial Superintelligence (ASI), a speculative AI that would surpass human intelligence in all aspects.

Artificial Narrow Intelligence (ANI)

(also called weak AI) refers to AI systems that are highly specialized and can perform only specific tasks or solve particular problems. Almost all current AI falls into this category of narrow intelligence rather than general, human-like problem-solving ability.

Artificial Superintelligence (ASI)

A hypothetical level of AI that far exceeds human intelligence and capabilities in all fields. ASI is a speculative concept often discussed in the context of AI's future, envisioning machines that not only match but vastly outperform the human mind, with unpredictable and possibly profound impacts on society if it were ever achieved.

Algorithm

An algorithm is a step-by-step procedure or set of rules for solving a problem or performing a task. In computing and AI, algorithms define how data is processed; for example, a search algorithm might outline how to retrieve information, or a learning algorithm might specify how an AI model adjusts its parameters during training. Well-designed algorithms enable computers to execute complex operations reliably and efficiently.

AI Alignment

Often called value alignment, this refers to the challenge of designing AI systems whose goals and behaviors are aligned with human values and intentions. In other words, an aligned AI will act in ways that are beneficial to humans and consistent with our ethical principles, rather than pursuing undesirable or harmful outcomes. Achieving proper alignment is especially crucial when developing very advanced or autonomous AI, to ensure these systems remain safe and beneficial.

Annotation (Data Labeling)

In the context of AI, annotation is the process of labeling or tagging data with additional information to make it understandable for machine learning models. For example, annotators might draw boxes around objects in photos or mark the correct answer in text, creating labeled datasets that supervised learning algorithms can learn from. High-quality annotations are crucial for training accurate models in tasks like image recognition and natural language processing.

Anthropomorphism

Attributing human traits, emotions, or intentions to non-human entities, including AI systems. For instance, users might anthropomorphize a chatbot by assuming it feels emotions or has a personality, even though it's actually a statistical model. While this habit can make interactions with AI feel more natural, it can also lead to misunderstandings about an AI system's true capabilities and limitations.

Augmentation (Human-AI Augmentation)

Rather than replacing humans, augmentation refers to using AI to enhance human abilities and productivity. In this model, AI systems handle repetitive or data-intensive sub-tasks (such as scanning large databases or summarizing documents), which frees up humans to focus on complex, creative, or interpersonal aspects of work. The goal is a collaborative dynamic where AI tools act as "copilots," boosting human decision-making and performance instead of working in isolation.

Automation

The use of technology (often including AI) to perform tasks with minimal human intervention. Automation ranges from simple scripts that automatically sort emails, to robotic process automation in business workflows, up to complex systems like self-driving cars. By letting machines handle repetitive or routine tasks, automation can increase efficiency and consistency – though it also raises discussions about workforce impacts and the importance of human oversight for critical systems.

AI Ethics

The moral principles and practices that guide the design, development, and deployment of AI in a responsible way. AI ethics addresses issues like fairness (avoiding biased or discriminatory outcomes), transparency (making it clear how decisions are made), privacy (protecting personal or sensitive data), and accountability (ensuring humans remain responsible for AI-driven outcomes). As AI systems become more widespread, ethical guidelines help mitigate harm and ensure AI technologies are developed for the common good.

AI Safety

Efforts and techniques aimed at preventing AI systems from causing unintended harm or behaving in undesirable ways. This field includes research into aligning AI objectives with human values, implementing safeguards or "guardrails" to catch and correct errors, and rigorously testing AI, especially in high-stakes areas like healthcare, finance, or transportation. AI safety becomes increasingly important as AI systems gain autonomy and complexity, to ensure they remain under control and act as intended.

AI Agent (Autonomous Agent)

A software entity that perceives its environment and takes actions autonomously in order to achieve specified goals. Advanced AI agents may use large language models or other AI techniques to analyze situations, make decisions, and carry out tasks with little to no human guidance. Examples include AI assistants that can plan travel itineraries based on a single request, or experimental agentic AIs that loop through tasks (like browsing the web and writing code) to fulfill an objective. The term "agent" highlights the AI system's ability to act on behalf of a user or system, initiating tasks and adapting to feedback along the way.

B

Bias (AI Bias)

In AI, bias refers to systematic errors or unfair prejudices in a model's outputs, caused by biased training data or flawed algorithms. For example, if an image recognition AI is trained mostly on photos of light-skinned people, it might perform poorly on recognizing people with darker skin – a bias reflecting its dataset. Biased AI can lead to discriminatory outcomes, so researchers prioritize techniques to detect and mitigate bias (such as balancing training data, or revising algorithms) in order to build fairer, more trustworthy AI systems.

Backpropagation

A fundamental algorithm for training artificial neural networks. Backpropagation works by calculating how much each connection weight in the network contributed to the error in the output, and then adjusting those weights in the opposite direction of the error (hence "back-propagating" the error). This iterative process of error calculation → weight adjustment allows the network to learn from mistakes on training examples. Backpropagation, combined with optimization methods like gradient descent, enabled the modern deep learning revolution by making it feasible to efficiently train very deep, multi-layered networks.

Big Data

A term for extremely large and complex datasets that are difficult to process and analyze using traditional methods. Big data often exhibits the "3 Vs": high volume (massive size), high variety (diverse types of data such as text, images, sensor logs), and high velocity (data being generated rapidly). In AI, big data is both fuel and a challenge – it can improve model performance by providing more examples to learn from, but it also requires robust infrastructure and algorithms to handle. Technologies like distributed computing and specialized storage have emerged to manage big data, enabling modern AI systems to crunch petabytes of information.

Black Box (AI Black Box)

A term describing an AI model (often a deep learning neural network) whose inner workings are not interpretable or transparent to humans. We can see what goes into the model (input) and what comes out (output), but the complex computations in between are opaque – even to the model's developers. For instance, a neural network might accurately flag loan applicants as high-risk or low-risk, but not provide a human-understandable reason. This lack of transparency is known as the "black box" problem and raises trust and accountability concerns. It has spurred interest in Explainable AI methods that can shine light on why black-box models make the decisions they do.

C

Chatbot

A software application that conducts a conversation via text or voice, often leveraging AI to understand queries and provide relevant responses. Some chatbots are simple scripts with pre-written responses, but more advanced conversational AI chatbots use natural language processing to interpret user input and generate fluid, context-appropriate replies. Chatbots are used for customer service (answering questions or troubleshooting), personal assistants, and entertainment. A well-known example is a customer support chatbot on a website that can help users reset a password or track an order through a dialog interface.

ChatGPT

An advanced AI chatbot developed by OpenAI that interacts in a conversational way and can answer questions, write text, or solve problems. ChatGPT is powered by a large language model (OpenAI's GPT series) and was initially built on GPT-3.5, with later versions using GPT-4. It gained widespread attention for its ability to produce human-like, coherent responses across a wide range of topics. Users can prompt ChatGPT with instructions or questions in plain language, and it will attempt to generate a helpful reply. The success of ChatGPT in late 2022 popularized AI assistants and led to a surge of interest in generative AI for everyday tasks.

Computer Vision

A field of AI focused on enabling computers to interpret and understand visual information from the world, such as digital images or videos. Tasks in computer vision include image classification (e.g. recognizing an image contains a cat), object detection (finding and labeling multiple objects in an image), facial recognition, image segmentation (outlining specific regions like tumors in a medical scan), and even generating images. By analyzing pixel patterns and features, computer vision systems can achieve outcomes like diagnosing diseases from X-rays, powering the vision of autonomous vehicles (identifying pedestrians, signs, other cars), or enabling your smartphone camera to automatically focus on faces.

Convolutional Neural Network (CNN)

A type of deep neural network architecture especially well-suited for processing grid-like data such as images. A CNN is composed of layers that apply many small filters (convolutions) to the input image, detecting simple features like edges at early layers and more complex features (like eyes or wheels) in deeper layers. These networks automatically and adaptively learn which visual features are important for a task (like recognizing a dog vs. a cat). CNNs have driven huge breakthroughs in image recognition and computer vision by dramatically improving accuracy in tasks like object detection and facial recognition.

D

Dataset

A collection of data, typically organized for a specific purpose, which is used for training or evaluating AI models. For instance, a dataset might consist of thousands of labeled images of cats and dogs for a model to learn to tell them apart. Datasets are often split into subsets: a training set (to learn from), a validation set (to tune model parameters and prevent overfitting), and a test set (to evaluate how well the trained model generalizes to new data). The quality and representativeness of a dataset are crucial – biased or insufficient data can lead to biased or inaccurate AI models.

Deep Learning

A subfield of machine learning that uses multi-layered neural networks to learn complex patterns from large amounts of data. "Deep" refers to the many layers in these neural networks – as data passes through each layer, the network can build increasingly abstract and high-level representations (for example, from raw pixels, a deep learning model might first detect edges, then shapes, then whole objects). Deep learning has achieved remarkable, human-level performance in tasks like image classification, speech recognition, and language translation by learning these representations directly from data. Its rise over the last decade was fueled by improvements in algorithms (like backpropagation), more powerful hardware (GPUs), and the availability of massive datasets (big data).

Diffusion Model

A type of generative model, especially used in image and audio generation, that learns to create new data by incrementally removing noise from random noise input. During training, a diffusion model takes training data (say, images) and progressively adds random noise until the data become pure noise; it learns how to partly reverse this process by predicting and reducing noise. At generation time, the model starts with a random noise and then iteratively refines it – essentially "denoising" step by step – until a coherent image or signal emerges. Diffusion models are behind state-of-the-art generative AI tools (e.g. Stable Diffusion for image synthesis) and are praised for producing high-quality, detailed outputs.

E

Ethics in AI

See AI Ethics for the principles guiding responsible AI design and use.

Explainable AI (XAI)

Techniques and methods that make an AI model's decisions and predictions understandable to humans. Because complex models like deep neural networks often operate as "black boxes," XAI seeks to provide insights into how the model is reasoning – for example, by highlighting which input features were most influential in a decision, or by simplifying the model's logic into human-readable rules. Explainability is especially important in domains like healthcare, finance, or law, where practitioners need to justify decisions. By increasing transparency, XAI helps build trust in AI systems and can assist developers in identifying flaws or biases in models.

Embedding (Vector Embedding)

A numerical representation of data – such as a word, sentence, or image – in the form of a vector (a list of numbers) that captures its meaning or features. The idea is to encode items so that those with similar meaning or content have embeddings that are close together in this vector space. For example, in an NLP embedding space, the words "king" and "queen" might end up near each other, while "king" and "banana" would be far apart. Embeddings enable semantic comparisons using math: an AI can compute distances between these vectors to find related concepts or retrieve relevant items. They are fundamental to tasks like semantic search, recommendation systems, and the way large language models represent knowledge internally.

F

Few-Shot Learning

A machine learning approach where a model can generalize and perform a new task from only a very small number of examples (sometimes just a handful). In few-shot learning, instead of needing thousands of labeled examples to learn a concept, the model leverages prior knowledge and sophisticated generalization to learn from, say, 5 or even 1 example per class. This is especially useful when data is scarce or costly to obtain – for instance, learning to detect a newly discovered rare disease from only a few medical scans. Techniques to achieve few-shot learning include meta-learning (learning how to learn) and using powerful pre-trained models that can adapt quickly to new data.

Fine-Tuning

The process of taking a pre-trained AI model and further training it on a smaller, task-specific dataset to adapt it to a particular application. Because the model has already learned general patterns from a large corpus of data (during pre-training), fine-tuning can teach it the nuances of a new task more efficiently than training from scratch. For example, one might fine-tune a general language model on medical transcripts to make it better at answering healthcare questions. Fine-tuning typically requires less data and time, and it leverages the investment in computing that went into the initial training of the foundation model.

Foundation Model

A large AI model (often based on deep neural networks) trained on a broad dataset at scale, which can be adapted to a wide range of downstream tasks. The term has emerged to describe models like OpenAI's GPT-3 or Google's BERT (in language) or CLIP (in vision) that serve as a common base – or foundation – for many applications. These models capture extensive knowledge about language or the world from their massive pre-training process, and developers can then build on them via fine-tuning or prompting for specific purposes. Foundation models have driven recent AI advancements because they can be reused for myriad tasks, reducing the need to train new models from scratch for each task.

G

Generative Adversarial Network (GAN)

A class of neural network architectures used for generative modeling, involving two networks in competition. One network, the generator, tries to create realistic fake data (e.g. generating a photo that looks like a real face), while the second network, the discriminator, tries to detect which data is real and which is generated. Through many training rounds, the generator gets better at fooling the discriminator, and the discriminator gets better at spotting fakes, until the generator's outputs are highly convincing. GANs have been used to create deepfake videos, generate artwork or design ideas, and even upsample low-resolution images – all by learning to mimic the patterns of real data.

Generative AI

Any AI technique that creates new data or content resembling the data it was trained on. Generative AI can produce text, images, music, program code, and more. Notable examples include language models like GPT-3 which can write coherent paragraphs or answer questions, and image models like DALL-E or Stable Diffusion which can generate original pictures from text prompts. These models work by learning the distribution of their training data (for instance, how words tend to follow each other, or how pixels form meaningful images) and then sampling from that distribution to create novel outputs. Generative AI opens up creative and practical applications, from assisting writers and artists to automating the creation of synthetic data for simulations.

GPU (Graphics Processing Unit)

A specialized processor originally designed for rendering graphics, now widely used to accelerate AI computations due to its ability to perform many operations in parallel. Training and running modern AI models – especially deep learning networks – involve heavy linear algebra (matrix and vector operations) on large datasets, which GPUs can handle much faster than traditional CPUs by splitting the work across hundreds or thousands of cores. The widespread availability of GPUs (and related hardware like TPUs) in the 2010s made it feasible to train large neural networks, and this hardware power was a key factor in the recent AI boom.

GPT (Generative Pre-trained Transformer)

A family of large language models developed by OpenAI, with notable versions including GPT-3 and GPT-4. Generative Pre-trained Transformer refers to the model's approach: it is a Transformer-based neural network that is first pre-trained on a vast amount of text data and is generative (able to produce text). GPT-3, for example, has about 175 billion parameters and was trained on hundreds of billions of words, enabling it to perform tasks from translation to essay writing. GPT-4 (introduced in 2023) is even more advanced and is multimodal, meaning it can accept images as part of its input and not just text. These GPT models are the engines behind conversational AI like ChatGPT and have demonstrated unprecedented versatility in language understanding and generation.

H

Hallucination (AI Hallucination)

The AI industry's term for when a model makes stuff up – i.e. it generates information that is incorrect or nonexistent, but presented as if it were factual. For example, a language model might confidently state a wrong historical date or invent a bibliographic reference that looks real but isn't. Hallucinations are a major issue for generative AI, as they can be misleading or even dangerous (imagine a medical AI giving a made-up treatment). These errors stem from the model's training limitations – it learns to produce plausible-sounding text, not to verify truth. Reducing AI hallucinations is an active area of research, involving better training data, model architecture tweaks, and allowing the AI to check its answers against reliable sources.

I

Inference

In AI, inference is the process of running a trained model on new data to generate outputs or predictions. It's essentially the "application" phase of machine learning – after an AI model has learned from examples during training, inference is when it uses that learned knowledge to do work in the real world. For instance, using a trained model to recognize speech in a live phone call, or to predict tomorrow's weather given today's data, are inference tasks. Inference can happen on powerful servers in the cloud or on devices like smartphones or IoT sensors at the edge. Often there's a focus on optimizing inference for speed and efficiency (e.g. using special hardware or model compression) so that AI-powered features feel instantaneous to users.

L

Large Language Model (LLM)

A type of AI model trained on a massive corpus of text to predict and generate language, effectively learning the statistical patterns of human language. LLMs (like OpenAI's GPT-3 or Google's PaLM) are built using deep neural network architectures – typically Transformers – and have billions of parameters. This large scale gives them a broad knowledge of grammar, facts, and even some reasoning ability acquired from the text they've seen. When you prompt an LLM with some input text or a question, it generates a continuation or answer by predicting the most likely next words based on its training. LLMs can perform a wide array of language tasks (translation, summarization, coding, Q&A, etc.) often with little or no task-specific training, which is why they underpin so many recent AI applications.

M

Machine Learning (ML)

A subset of AI that focuses on algorithms enabling computers to learn from data and improve their performance on tasks without being explicitly programmed for each case. In traditional programming, humans write rules; in machine learning, the system discovers rules and patterns from example data. There are several types of ML: in supervised learning, the algorithm learns from labeled examples (e.g. images with labels "cat" or "dog"); in unsupervised learning, it finds structures in unlabeled data (like grouping similar customers together); and in reinforcement learning, it learns by trial and error with rewards. ML techniques power everything from recommendation engines and spam filters to self-driving car vision systems, by continually adjusting models as more data becomes available.

Model (AI Model)

In AI, a model refers to the mathematical or computational representation of a learned pattern or task. It's essentially the end result of the training process – the model encapsulates what the algorithm has learned from the training data. For example, a trained model could take an email as input and output a prediction "spam" or "not spam" based on patterns it learned. Internally, a model consists of parameters (numbers) and structure (e.g. the layers of a neural network) that together define how input data is transformed into output. Once a model is trained and validated, it can be deployed to make predictions on new data (this use phase is called inference). Models can range from simple linear regressions to giant neural networks with billions of parameters.

Multimodal AI

AI systems or models that can process and relate information from multiple modalities (types of data), such as text, images, audio, video, etc.. For example, a multimodal AI might take an image and a question about that image as input, and produce a text answer – it has to combine visual understanding with language understanding to respond correctly. By integrating multiple data sources, multimodal models aim for a more holistic understanding of context (since humans, for instance, use both sight and sound together). A recent development in this area is multimodal large language models like GPT-4, which can accept both images and text in their input. Such models can describe an image you upload, answer questions about it, or use the image as additional context for a conversation, demonstrating a step toward AI that "sees and talks."

N

Neural Network (Artificial Neural Network)

A computational model inspired by the human brain's network of neurons, consisting of layers of interconnected nodes (neurons) that transform input data to recognize patterns and make predictions. Each connection in a neural network has a weight – a parameter that amplifies or dampens the signal – and the network learns by adjusting these weights based on errors (using algorithms like backpropagation). Simple neural networks have one or two layers (sometimes called "perceptrons"), while deep neural networks have many layers and can model very complex relationships. Neural networks come in various architectures specialized for different data: e.g. convolutional neural networks for images, recurrent neural networks for sequential data like text or time series, and transformers for language and beyond. They are the cornerstone of most modern AI systems, capable of learning functions that were impractical to hard-code.

Natural Language Processing (NLP)

A field of AI that gives computers the ability to understand, interpret, and generate human language. NLP encompasses both text and spoken language. Common NLP tasks include machine translation (e.g. translating French to English), sentiment analysis (determining if a sentence sounds positive or negative in tone), named entity recognition (finding names of people/places in text), speech recognition (converting audio speech to text), and text generation. NLP techniques combine linguistic rules with machine learning algorithms – increasingly dominated by large neural network models – to handle the complexity and ambiguity of language. Thanks to NLP, we have chatbots that can hold conversations, email filters that catch spam because of phrasing, and voice assistants that follow spoken commands.

O

Overfitting

A common pitfall in machine learning where a model learns the training data too well – including its noise or random fluctuations – and as a result performs poorly on new, unseen data. An overfitted model has essentially memorized the training examples instead of generalizing from them. For instance, imagine a classifier that perfectly classifies every training image (even the oddities), but then fails on slightly different images; it's tuned to the training set's quirks. Overfitting often happens when a model is very complex relative to the amount of training data. To prevent it, practitioners use techniques like cross-validation, regularization (adding penalties for complexity), and simply gathering more data. The opposite problem is underfitting, where a model is too simple to capture the underlying pattern and performs poorly even on training data.

P

Parameter (Model Parameter)

A variable in an AI model that is learned from training data, which defines how the model processes input and generates output. In a neural network, for example, the weights of connections and biases are parameters – each influences the signal flowing through the network. During training, the learning algorithm adjusts these parameters to minimize errors. The number of parameters in a model (often measured in millions or billions) is one way to gauge its size or capacity; for instance, GPT-3 famously has 175 billion parameters. The trained values of all parameters collectively determine the model's behavior. (By contrast, hyperparameters are settings chosen by the developer, like the learning rate or number of layers, not learned from data.)

Pre-Training

The initial training phase where a model is trained on a broad, general task or dataset as a foundation, before being fine-tuned for a specific application. In the context of language models, pre-training might involve learning to predict the next word in billions of sentences from the internet (an unsupervised task), which teaches the model grammar, facts, and reasoning abilities. After pre-training, the model can be fine-tuned with a much smaller set of task-specific data (like a set of legal documents, if we want a legal assistant AI) to specialize it. Pre-training is powerful because it allows the model to reuse general knowledge for many different tasks, significantly reducing the data and time needed for each new task.

Prompt

The input or instruction given to a generative AI model to produce a response. A prompt can be as simple as a single question ("What is the capital of France?") or a complex directive ("Write a short story about a detective solving a case at a bakery, in the style of Sherlock Holmes."). The model processes the prompt and then generates an answer or continuation based on it. Crafting a clear and precise prompt is important because the AI's output heavily depends on how the request is phrased and what context is given. As users have discovered, rephrasing or adding detail to prompts can lead to significantly better results from models like ChatGPT.

Prompt Engineering

The practice of designing and refining prompts to guide AI models (especially large language models) toward producing the desired output. Because these models can follow a wide range of instructions, prompt engineering involves finding the right way to ask a question or give instructions to yield the best response. This might mean providing the AI with step-by-step bullet points to follow, giving examples of the kind of answer you want (few-shot prompting), or adding constraints like "answer in 100 words or fewer." Prompt engineering has become an important skill for AI practitioners and users, as it can significantly improve an AI's performance on a task without changing the model itself – only the way it's queried.

R

Reinforcement Learning (RL)

An approach to AI where an agent learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties. The agent's objective is to learn a policy (a strategy mapping situations to actions) that maximizes the cumulative reward over time. This trial-and-error learning paradigm is inspired by behavioral psychology (think of training a pet: good behavior is rewarded, bad behavior might get no reward or a timeout). Reinforcement learning has led to impressive results, such as AIs that learn to play Atari games or Go at superhuman levels, and is used in robotics, game theory, and scenarios where long-term planning is needed. Key concepts include the state (the current situation), action (what the agent can do), reward (feedback signal), and Q-values or policies that the agent develops.

Reinforcement Learning from Human Feedback (RLHF)

A technique for aligning AI behavior with human preferences by incorporating human feedback into the training loop. In RLHF, humans first evaluate or rank a variety of outputs that an AI model produces (for example, different answers a chatbot might give to the same question). Those human preferences are used to train a reward model that predicts what humans would prefer. Then, using reinforcement learning, the AI model is further trained to maximize the reward according to that learned reward model – effectively tuning the AI to produce outputs that humans find more helpful or appropriate. RLHF was instrumental in training models like ChatGPT to follow instructions and avoid undesirable outputs, as it curbs the model's tendencies by teaching it human-endorsed behavior.

Robotics

An interdisciplinary field (combining engineering, computer science, and AI) that involves the design, construction, and operation of robots. Robots are machines – often programmable and sometimes autonomous – that can perform tasks and interact with the physical world. AI plays a significant role in modern robotics by giving robots perception (e.g. using computer vision to "see"), decision-making abilities, and adaptability. For example, AI algorithms help self-driving cars (considered robots) recognize lanes and pedestrians, or enable a warehouse robot to learn the most efficient way to sort packages. Robotics covers a spectrum from industrial robots (fixed machines on assembly lines) to social robots (like helper robots or drones) – and as AI advances, robots are becoming more capable and versatile in unstructured environments.

S

Supervised Learning

A machine learning paradigm where the model is trained on labeled data – each training example includes an input and a correct output (label) that the model should learn to produce. The learning process is "supervised" by these labels, as the model's predictions can be directly compared to the known answers and adjusted accordingly. For instance, a supervised learning system could learn to detect spam emails by training on a dataset of emails labeled as "spam" or "not spam." Over time, the model tunes itself to accurately map inputs to outputs. This approach is very powerful when lots of labeled data is available and is used in tasks like image recognition, speech-to-text, and medical diagnosis (with labeled medical images or records).

Unsupervised Learning

A machine learning approach where the model is given data without explicit labels and must find structure or patterns on its own. The model might cluster similar data points together, discover hidden correlations, or reduce data to its most important dimensions. For example, an unsupervised algorithm could analyze millions of customer transactions and automatically group customers into segments with similar purchasing habits, without being told what those segments are. Common unsupervised techniques include clustering algorithms (like k-means), principal component analysis (PCA) for dimensionality reduction, and autoencoders. Unsupervised learning is useful for exploratory data analysis, compression, anomaly detection (finding outliers that don't fit any pattern), and as a pre-training step for other types of learning.

Symbolic AI (GOFAI)

An approach to AI, dominant in the early decades of AI research, that uses explicit symbolic representations and logic to solve problems. In symbolic AI, knowledge is manually encoded in symbols (like words or logical propositions) and manipulated with rules – for example, an expert system containing IF-THEN rules to diagnose illnesses based on symptoms. Also known as "Good Old-Fashioned AI," this approach excels at transparency (it's clear why the system reached a conclusion, since one can trace the logical rules) and works well for problems that can be neatly described by rules. However, it struggles with the ambiguity and noise of real-world data (e.g. understanding natural images or language). Modern AI often blends symbolic reasoning with machine learning (which is data-driven) to get benefits of both – symbolic AI for reasoning and memory, and statistical AI for perception and pattern recognition.

Technological Singularity

A theoretical future point at which technological growth, particularly in AI, becomes uncontrollable or exponential, resulting in unfathomable changes to human civilization. Often the singularity is imagined to occur when AI attains the capability to improve itself recursively – creating ever smarter AIs without human intervention – quickly surpassing human intelligence (an analogy is how an upgraded computer can design its next, even more powerful upgrade). The term is associated with futurist Ray Kurzweil and others. While some see the singularity as a moment of great promise (solving aging, energy, etc.), others worry about existential risks, as humanity might lose the ability to predict or direct what superintelligent AI will do after that point.

Speech Recognition (Automatic Speech Recognition, ASR)

Technology that enables a computer to convert spoken language into text. Modern speech recognition systems use advanced AI models (often deep learning networks) trained on thousands of hours of audio to handle variations in accent, pronunciation, and noise. ASR is what allows virtual assistants like Siri, Alexa, or Google Assistant to understand your spoken questions and commands. It's also used in dictation software, live transcription services (like captioning a live video call), and voice control interfaces in cars or smart devices. While far from perfect, speech recognition has improved dramatically with AI, achieving high accuracy for many languages and contexts, and continues to expand access and usability in technology through hands-free, voice-driven interaction.

T

Transformer

A groundbreaking neural network architecture introduced in 2017 that uses a mechanism called self-attention to efficiently process sequences of data (such as words in a sentence) in parallel. Unlike older sequence models (e.g. RNNs) that processed words one after another, transformers can look at all the words (or other elements) at once and learn which ones are important in relation to others, regardless of their position. This ability to capture long-range dependencies and the meaning of words in context made transformers especially effective for natural language tasks. Transformers are the backbone of most state-of-the-art language models and have also been applied in vision (Vision Transformers) and other domains. In short, the Transformer architecture enabled the current leap in AI by allowing models to be trained on unprecedented amounts of data and context, with much better performance and efficiency than previous approaches.

Training (Model Training)

The process by which an AI model is fed data and learns from it, by adjusting internal parameters to improve its performance on a task. During training, the model makes predictions on the input data and an optimizer algorithm tweaks the model's parameters to reduce the difference between predictions and the correct answers. This usually happens iteratively: the training dataset is passed through the model multiple times (in epochs), and the model gradually improves. Training is typically computationally intensive – often done on GPUs or TPUs – and can take anywhere from seconds (for a tiny model) to weeks or months (for very large models on huge datasets). Once training is complete, the resulting model can be used in inference mode to make predictions on new data. A crucial part of training is preventing overfitting (so the model generalizes well) and validating progress using techniques like a held-out validation set.

Transfer Learning

A machine learning technique where knowledge gained in training one model on a particular task or domain is re-used to jump-start training on another task or domain. Rather than starting from scratch, developers take a model that's already been trained on a large, general dataset (for example, an image recognition model trained on millions of everyday images) and then fine-tune it on a smaller, specific dataset (say, medical images). The pre-trained model's earlier layers already detect general patterns (edges, textures), which remain useful, and only the later layers adapt to the new task. Transfer learning can dramatically reduce the amount of data and time needed for the new task and often results in better performance. It's widespread in NLP and vision – for instance, many language models today use a large pre-trained model and then specialize it via fine-tuning for tasks like sentiment analysis or question answering.

Turing Test

A classic experiment proposed by Alan Turing in 1950 as a criterion for whether a machine can exhibit human-like intelligence. In a Turing Test setup, a human evaluator engages in a natural language conversation with two hidden entities – one a human and one a machine – and tries to determine which is which. If the machine can convince the evaluator that it's human (typically by giving answers indistinguishable from a person's), it's said to have passed the test. The Turing Test doesn't directly test for "consciousness" or understanding; it's purely behavioral. Over the years, a few chatbots have claimed to fool judges in restricted settings, but no AI has universally passed as human under rigorous conditions. The test remains a touchstone in AI discussions about language ability and has inspired many competitions, but it's also been critiqued for being too focused on deception rather than true intelligence or usefulness.

Zero-Shot Learning

A machine learning capability where a model successfully handles tasks or recognizes classes that it wasn't explicitly trained on, by leveraging auxiliary information or general knowledge. For example, an image model that has learned about animals might recognize a zebra even if it never saw one during training, if it knows a zebra is "like a horse with stripes" – the model uses its knowledge of horses plus the concept of stripes to bridge the gap. In NLP, zero-shot learning might mean a language model can perform a task (say, translation or sentiment analysis) just by being instructed in plain language, despite not being specifically trained for that task. Zero-shot capability is a hallmark of very general models and is closely tied to the versatility we see in large pre-trained models (they can often generalize in surprising ways when prompted correctly).