A dictionary of AI terms with some fundamental concepts to avoid getting lost in the tide of a revolutionary technology.

Artificial intelligence is already an active part of many of our activities. Although we use it, we often don't really know how it works. Below you will find some concepts that can help you understand how it works.

Setting instructions

Instruction tuning is an approach where a pre-trained model is adapted to perform specific tasks by providing a set of guidelines or directives that describe the desired operation.

Fine-tuning

The process of adapting a pre-trained model to a specific task by training it on a smaller data set. For example, an image classification model trained on all intersection photos can be refined to detect when a car runs a red light.

Hallucination

Hallucination refers to a situation where an AI system, especially in cases where it handles natural language processing, begins to generate results that may be irrelevant, meaningless or incorrect based on the input provided. Frequently, this occurs when the AI system is unsure of the context, relies too heavily on its training data, or lacks an adequate understanding of the subject matter presented to it.

Anchor

Anchoring is the process of anchoring AI systems to real-world experiences, knowledge or data. The goal is to improve AI's understanding of the world, so that it can effectively interpret and respond to user inputs, queries and tasks. Anchoring helps AI systems become more context-aware, enabling them to provide more appropriate, relatable and relevant responses or actions.

Stacking

Stacking is a technique in AI that combines multiple algorithms to improve overall performance. By combining the strengths of multiple AI models, stacking compensates for the weaknesses of each model and achieves more accurate and robust output in a variety of applications, such as image recognition and natural language processing.

Machine learning (machine learning)

A subfield of AI that encompasses the development of algorithms and statistical models that allow machines to improve their performance with experience. An example would be a machine learning algorithm that could predict the likelihood of customer loss based on past behaviour.

Collective learning

Collective learning is an AI training approach that leverages diverse skills and knowledge across multiple models to achieve more powerful and robust intelligence.

Zero-trigger learning or zero training

Zero-trigger learning is a technique in which a machine learning model can recognise and classify new concepts without any labelled examples.

Non-supervised learning

A type of machine learning in which a model is trained on unlabelled data to find patterns or features in the data. Example: An unsupervised learning algorithm that can cluster similar images of handwritten digits based on their visual characteristics.

Enhanced learning

A type of machine learning in which a model learns to make decisions by interacting with its environment and receiving feedback through rewards or penalties. GPT uses enhanced learning from human feedback. When tuning GPT-3, human annotators provided examples of the model's desired behaviour and classified the model's outputs

Deep learning

A subfield of machine learning that uses multi-layered neural networks to learn from data. An example would be, in this case, a deep learning model that could recognise objects in an image by processing it through multiple layers of neural networks.

Automation

Automation refers to the use of technology to perform various tasks with minimal human intervention.

Capability of control

Controllability refers to understanding, regulating and managing the decision-making process of an AI system, ensuring its accuracy, safety and ethical behaviour, and minimising the potential for unintended consequences.

Chatbot

An easy-to-use interface that allows the user to ask questions and receive answers. Depending on the back-end system powering the chatbot, it can be as basic as prescribed answers to fully conversational AI that automates problem solving.

ChatGPT

A chat interface built on top of GPT-3.5. GPT-3.5 is a large language model developed by OpenAI that is trained on a massive amount of text data from the internet and optimised to perform a wide range of natural language tasks. Example: GPT-3.5 has been refined for tasks such as language translation, text summarisation and question answering.

Validation Data

A subset of the dataset used in machine learning that is separate from the training and test datasets. It is used to fit the hyperparameters (i.e., the architecture, not the weights) of a model.

Unstructured Data

Unstructured data is any information that is not organised in a predefined model or structure, making it difficult to collect, process and analyse.

Model linking

Model chaining is a technique in data science where several machine learning models are linked in sequence to make predictions or analyses.

Data enrichment

Data enrichment is a technique used to artificially increase the size and diversity of a training set by creating modified copies of existing data. It involves making minor changes such as flipping, resizing or adjusting the brightness of images, to improve the dataset and prevent over-fitting of models.

Latent space

In machine learning, this term refers to the compressed representation of data that a model (such as a neural network) creates. Similar data points are closer together in latent space.

Extensibilidad

Extensibility in AI refers to the ability of AI systems to expand their capabilities to new domains, tasks and datasets without the need for complete retraining or major architectural changes.

Extracción

Extraction is the ability of generative models to analyse large data sets and extract patterns, trends and specific pieces of relevant information.

GPT-3

GPT-3 is the third version of the GPT-n model series. It has 175 billion parameters (adjustable knobs) that can be weighted to make predictions. Chat-GPT uses GPT-3.5, which is another iteration of this model.

GPT-4

GPT-4 is the latest model addition to OpenAI's deep learning efforts and is a significant milestone in deep learning scalability. GPT-4 is also the first GPT model that is a large multimodal model, meaning it accepts image and text inputs and emits text outputs.

Weak-to-strong generalisation

Weak-to-strong generalisation is an AI training approach that uses less capable models to guide and constrain more powerful ones towards better generalisation beyond their narrow training data.

Language Large Models (LLM)

A type of deep learning model trained on a large dataset to perform natural language understanding and generation tasks. There are well-known models such as BERT, PaLM, GPT-2, GPT-3, GPT-3.5 and the innovative GPT-4. All these models vary in size (number of parameters that can be tuned), in the breadth of tasks (encoding, chat, scientific, etc.), and in the knowledge they have been trained on.

Inference

The process of making predictions with a trained machine learning model, i.e. putting into practice what the AI has learned during its training.

Engineering prompts

Identification of inputs (prompts) that result in meaningful outputs. So far, prompts engineering is essential for LLMs. LLMs are a fusion of layers of algorithms that consequently have limited controllability with few opportunities to intervene and override behaviour. A use case for prompt engineering would be to create a collection of templates and wizards to control a commercial copywriting application.

Artificial intelligence (AI)

The simulation of human intelligence in machines that are programmed to think and learn like humans. Example: an autonomous car that can navigate and make decisions on its own using AI technology.

Artificial Intelligence General (AGI)

General Artificial Intelligence (GAI) refers to an AI system that possesses a wide range of cognitive abilities, much like humans, allowing it to learn, reason, adapt to new situations and develop creative solutions across diverse tasks and domains, rather than being limited to specific tasks as narrow AI systems are.

Conversational AI

A branch of AI development focused on developing systems that can understand and generate human-like language and carry on a back-and-forth conversation. An example of this would be a chatbot that could understand and respond to customer queries in a similar way to a conversation between people.

Generative AI

Generative AI models create new data or content by discovering patterns in data inputs or training data. For example, creating an original short story based on analysis of existing published short stories.

Associative Memory

Associative memory refers to the ability of a system to store, retrieve and process related information based on connections between elements, enabling it to efficiently identify and use relevant data for decision making.

Sequence modelling

A subfield of natural language processing that focuses on modelling sequential data such as text, speech or time series data. In this case, a sequence model would be one that can predict the next word in a sentence or generate coherent text.

Base model

Base models are a broad category of AI models that include large language models and other types of models such as computer vision and reinforcement learning models. They are called "base" models because they serve as the foundation on which to build applications, catering to a wide range of domains and use cases.

Deterministic model

A deterministic model follows a specific set of rules and conditions to reach a definite outcome, operating on a cause-and-effect basis.

Discriminative model

Discriminative models are algorithms designed to directly model and learn the boundary between different classes or categories in a data set.

OpenAI

The company that developed ChatGPT. OpenAI is, broadly speaking, a research company that aims to responsibly develop and promote friendly AI. One example is its GPT-3 version of OpenAI, one of the largest and most powerful language models available for natural language processing tasks.

Optimisation

The process of adjusting the parameters of a model to minimise a loss function that measures the difference between model predictions and true values. This would be the case when you want to optimise the parameters of a neural network using a gradient descent algorithm to minimise the error between the model predictions and the true values.

Parameters

In machine learning, parameters are the internal variables that the model uses to make predictions. They are learned from the training data during the training process. As an example, in a neural network, weights and biases are parameters.

AI plugin

AI plugins are specialised software components that allow AI systems to interact with external applications and services.

Pre-training

The initial training phase of a machine learning model where the model learns general features, patterns and representations of the data without specific knowledge of the task to which it will later be applied. This unsupervised or partially supervised learning process allows the model to develop a fundamental understanding of the underlying distribution of the data and extract meaningful features that can be leveraged for subsequent fine-tuning on specific tasks.

Bias or bias

These are the assumptions made by an AI model about the data. A "bias-variance trade-off" is the balance that must be struck between the assumptions a model makes about the data and the amount that a model's predictions change, given different training data. Inductive bias is the set of assumptions a machine learning algorithm makes about the underlying distribution of the data.

Natural language processing

A subfield of AI that involves programming computers to process massive volumes of linguistic data. It focuses on transforming free-form text into a standardised structure.

Speech Processing

Speech processing in AI refers to the sequence of speech-to-text conversion followed by text-to-speech synthesis.

Recurring Prompts

Recurring prompts are a strategy for guiding AI models such as OpenAI's GPT-4 to produce higher quality results. It involves providing the model with a series of prompts or questions that build on previous answers, refining both the context and the AI's understanding to achieve the desired result.

Reasoning

AI reasoning is the process by which artificial intelligence systems solve problems, think critically and create new knowledge by analysing and processing available information, enabling them to make well-informed decisions in various tasks and domains.

Multi-hop reasoning

Multi-jump is a term often used in natural language processing, and more specifically in those tasks related to machine-reading comprehension. It refers to the process by which an AI model retrieves answers to questions by connecting multiple pieces of information present in a given text or across multiple sources and systems, rather than directly extracting information from a single passage.

Automatic Speech Recognition

Automatic Speech Recognition (ASR) is a technology that transcribes spoken language into text.

Neural Network

A machine learning model inspired by the structure and function of the human brain that is composed of layers of interconnected nodes or "neurons". They enable pattern recognition and problem solving, based on machine learning or deep learning, seeking to model the analysis mechanisms of the human brain.

Adversarial generative networks

Adversarial or antagonistic generative networks are a powerful example of a neural network, capable of generating new, never-before-seen information that closely resembles training data.

Backpropagation

This is an algorithm frequently used in the training of artificial neural networks. It adjusts the weights of the connections between neurons to minimise the error in the prediction of a model. It is effective for, for example, pattern recognition or image classification.

Singularity

In the context of AI, singularity (also called technological singularity) refers to a hypothetical future point in time when technological growth becomes uncontrollable and irreversible, leading to unpredictable changes in human civilisation.

Overfitting

A problem that occurs when a model is too complex, performing well on training data, but poorly on unseen data. An example of overfitting would be the case of a model that has memorised training data rather than learning general patterns, and therefore performs poorly on new data.

Stable Diffusion

Stable Diffusion is an open source artificial intelligence system that uses deep learning to generate images from natural textual cues.

Text-to-speech

Text-to-Speech (TTS) is a technology that converts written text into spoken voice output. It allows users to listen to written content being read aloud, typically using synthesised speech.

Tokenisation

The process of breaking text into individual words or sub-words for input into a language model. Example: tokenising a sentence "I am ChatGPT" into the words: "I," "am," "Chat," "G," and "PT."

Transformador (transformer)

This represents a type of neural network architecture designed to process sequential data, such as text. The transform architecture learns context and therefore meaning. It is used in models such as ChatGPT for natural language processing tasks.

Cross-validation

A method used to evaluate the results of an analysis and ensure the accuracy of the predictions of machine learning models.

Oracle Dedicated Region Solution Toolkit Overview