Glossary and FAQ of tems used

This page will be updated from time to time. Last update: 26th May

AI

Artificial intelligence (AI) is a branch of computer science that aims to create machines and systems that can perform tasks that normally require human intelligence and abilities. AI can be applied to various domains and problems, such as speech recognition, computer vision, natural language processing, robotics, gaming, healthcare, and more. AI systems can learn from data and experience, reason and plan, communicate and interact, and perceive and understand their environment.

AI can be classified into different types based on its capabilities and goals. Some common types are:

- Narrow AI: This type of AI is designed to perform a specific task or function, such as playing chess, recognizing faces, or recommending products. Narrow AI systems are often based on machine learning techniques that enable them to learn from data and improve their performance over time. However, they cannot generalise beyond their domain or task and lack common sense or general intelligence.

- General AI: This type of AI is the ultimate goal of AI research, as it aims to create machines that can perform any intellectual task that a human can do. General AI systems would have human-like intelligence and abilities, such as reasoning, learning, creativity, and self-awareness. However, this type of AI is still a theoretical concept and does not exist yet.

- Super AI: (or ASI) This type of AI is an extension of general AI, as it refers to machines that can surpass human intelligence and capabilities in all aspects. Super AI systems would have superior knowledge, skills, speed, memory, and creativity than humans. Some people believe that super AI could pose an existential threat to humanity if not aligned with human values and goals.

Using Consensus provides the definition of AI as:

A complex concept involving algorithms simulating human intelligence, socio-technological apparatuses, and social actor characteristics, with varying emphasis on technical functionality or human-like thinking depending on the context.

Alternatively

It can be argued that AI exists as a machine that can undertake interpolation, and extrapolation tasks, but exhibits no creativity, no consciousness, no dreaming, never has a sense of confidence, purpose or achievement.

Algorithms (for machine learning)

What is AGI?

AGI stands for Artificial General Intelligence, which is a type of artificial intelligence that can perform any intellectual task that a human or an animal can do. Unlike narrow AI, which is designed to solve specific problems, AGI aims to have general cognitive abilities that can adapt to any situation or goal. AGI is a major goal of some artificial intelligence research, but it has not been achieved yet. Some of the challenges of creating AGI include defining and measuring intelligence, replicating human common sense and creativity, and ensuring ethical and safe outcomes.

Buffer of Thoughts

The advantages of Buffer of Thought (BoT) to Tree-of-Thoughts (ToT) are discussed in the context of single-query and multi-query reasoning processes for large language models (LLMs). BoT and ToT are categorised as multi-query reasoning frameworks.  However, according to the paper 'Buffer of Thoughts: Thought-Augmented Reasoning with Large Language Models' BoT overcomes limitations seen in single and multi-query reasoning by using a meta-buffer, a library of high-level thoughts.

This meta-buffer contains a series of thought templates refined from numerous problem-solving techniques, and these thoughts are applicable across multiple tasks. Each problem uses a relevant thought template with a specialised reasoning structure, enabling efficient thought-augmented reasoning. BoT dynamically updates its meta-buffer using a buffer manager. BoT is more accurate and efficient than previous prompting methods and outperforms ToT in inference time. When evaluated against the Llama3-70B, BoT with the Llama3-8B shows the potential to surpass it. 

Cognitive blindness 

refers to a failure to perceive or understand information that is present due to various reasons. The reasons can include anxiety, lack of attention, or insufficient knowledge. There are different types of cognitive blindness, such as metacognitive blindness, which occurs when individuals fail to recognize their own cognitive processes. Cognitive blindness can also occur in relation to specific topics, such as AIDS awareness. Inattentional blindness is another type of cognitive blindness that occurs when people fail to notice unexpected objects when their attention is focused elsewhere. However, it is important to note that cognitive blindness is not always a permanent condition and can be addressed through interventions such as publicizing knowledge and building self-prevention awareness

GPT

Generative Pre-trained Transformer architecture. The model has been trained on a diverse range of internet text, allowing it to generate human-like text in response to prompts given to it. A useful video is below that explains much about training LLMs etc

Neural Network

A neural network is a computational model that simulates the functioning of biological neurons and their connections. Neural networks are composed of artificial neurons, also called nodes, that receive inputs from other nodes or external sources, process them using a mathematical function, and produce an output that can be transmitted to other nodes or used as a final result. Neural networks can learn from data and adjust their weights and biases accordingly, using various learning algorithms. Neural networks are widely used in artificial intelligence applications, such as speech recognition, image analysis, natural language processing, and adaptive control. Neural networks can perform complex tasks that are difficult to solve using conventional programming or statistical methods.

Machine Learning

Machine learning is a field of inquiry that studies how computer systems can learn from data and improve their performance on various tasks without being explicitly programmed. Machine learning is a subfield of artificial intelligence, which aims to create machines that can imitate intelligent human behaviour. Machine learning algorithms use mathematical models and statistical methods to analyse data, identify patterns, and make predictions or decisions. Machine learning algorithms can be classified into different types based on the nature of the data and the learning process, such as supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, and deep learning. Machine learning has many applications in diverse domains, such as medicine, computer vision, natural language processing, robotics, speech recognition, agriculture, and data mining. Machine learning is also closely related to other fields of study, such as computational statistics, mathematical optimization, and neural networks. Machine learning is an active and growing area of research and innovation that has the potential to transform many aspects of human society.

What is ChatGPT actually doing?

ChatGPT is a conversational AI model developed by OpenAI based on the Generative Pretrained Transformer 3 (GPT-3) architecture. The model has been trained on a diverse range of internet text, allowing it to generate human-like text in response to prompts given to it. ChatGPT can answer questions, converse on a variety of topics, and generate creative writing pieces. ChatGPT is designed to interact in a dialogue format, which makes it possible for it to answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests. ChatGPT is part of the broader field of artificial intelligence known as natural language processing (NLP), which seeks to teach computers to understand and interpret human language. ChatGPT is built using a deep learning architecture called the Transformer, which enables it to learn patterns in language and generate text that is coherent and human-like. ChatGPT is one of the most advanced language models available today and has the potential to revolutionise the way we interact with computers and digital systems. LLM A large language model (LLM) is a type of artificial intelligence (AI) algorithm that uses deep learning techniques and massively large data sets to understand, summarise, generate and predict new content. LLMs are trained on large quantities of unlabelled text using self-supervised learning, which means they learn from the data itself without human intervention or guidance. LLMs typically have billions of parameters, which are the variables that determine how the model processes the input and produces the output. LLMs use a neural network architecture called transformer, which enables them to capture long-range dependencies and complex relationships among words and sentences.

LLMs emerged around 2018 and have shown remarkable performance on a wide range of natural language processing (NLP) tasks, such as question answering, text summarization, text generation, sentiment analysis, machine translation and more. LLMs are also able to demonstrate general knowledge about the world and memorise facts from the data they are trained on. LLMs are sometimes referred to as foundation models, because they serve as the basis for further optimization and specialisation for specific domains and applications. However, LLMs also pose some challenges and risks, such as ethical, social and environmental implications. For example, LLMs may generate inaccurate, biassed or harmful content that can mislead or harm users or society. LLMs may also consume a lot of energy and resources during training and inference, which can have a negative impact on the environment. Moreover, LLMs may exhibit unpredictable or emergent behaviours that are not intended by the designers or users, such as hallucinations or abilities that were not explicitly programmed into the model. Therefore, LLMs require careful evaluation, monitoring and regulation to ensure their safe and beneficial use.

Mixture of Agents (MOA)

Mixture of Agents (MOA) is an approach that utilises the collective strengths of multiple large language models (LLMs) to enhance the quality of output. This approach surpasses the performance of leading frontier models like GPT 4.0 by a significant margin, achieving a score of 65.1 on alpaca eval 2.0, compared to GPT 4.0's 57.5. This is achieved through the collaboration of multiple open-source LLMs, rather than relying on a single, large, generalist model.

The MOA architecture consists of multiple layers, with each layer typically containing three agents. These agents collaborate to generate the final output for a given prompt. The outputs from each layer are used as auxiliary information by the agents in the subsequent layer, enabling a refinement of responses through the integration of diverse capabilities and insights from various models.

The research highlights the significance of "collaborativeness of LLMs," a phenomenon where an LLM generates better responses when presented with outputs from other models, even if those models are individually less capable.** This collaborative approach, using multiple models with different strengths and weaknesses, leads to more robust and versatile output.

The research identifies three key roles for the LLMs within the MOA framework:

Proposers: Generate initial reference responses, offering diverse perspectives for the aggregator.

Aggregators: Synthesise the responses from proposers into a single high-quality response.

Ranker: (Optional) An LLM can be used to select the best response from the aggregated options.

The research demonstrates that increasing the number of proposers generally leads to enhanced output quality, highlighting the value of diverse perspectives. However, while MOA achieves higher accuracy, it comes with a trade-off: slower inference time (time to first token). Reducing this latency is an area for future research.

NLP NLP stands for natural language processing, which is a branch of artificial intelligence that deals with the interaction between computers and human languages. NLP aims to enable computers to understand, analyse, generate and manipulate natural language texts or speech. Some of the applications of NLP include machine translation, speech recognition, sentiment analysis, information extraction, text summarization, question answering and chatbots. NLP is a challenging and multidisciplinary field that requires knowledge and skills from linguistics, computer science, mathematics and statistics. NLP involves various tasks and subfields, such as: - Tokenization: splitting a text into smaller units called tokens, such as words or punctuation marks. - Morphology: analysing the structure and formation of words, such as stems, prefixes and suffixes. - Syntax: analysing the grammatical structure and rules of sentences, such as parts of speech and dependency relations. - Semantics: analysing the meaning and logic of words and sentences, such as synonyms, antonyms and entailment. - Pragmatics: analysing the context and purpose of language use, such as speech acts and implicatures. - Discourse: analysing the structure and coherence of longer texts or conversations, such as paragraphs and dialogues. - Phonetics: analysing the sounds and pronunciation of speech, such as vowels and consonants. - Phonology: analysing the patterns and rules of sounds in a language, such as stress and intonation. - Prosody: analysing the rhythm and melody of speech, such as pitch and tone. NLP relies on various techniques and methods to perform these tasks and subfields, such as: - Rule-based systems: using predefined rules and dictionaries to process natural language based on its structure and grammar. - Statistical methods: using mathematical models and algorithms to learn from data and make predictions based on probabilities and frequencies. - Machine learning: using artificial neural networks and other learning algorithms to automatically learn from data and improve performance based on feedback. - Deep learning: using advanced neural networks with multiple layers and complex architectures to perform complex natural language tasks with high accuracy. NLP is a rapidly evolving and expanding field that has many current challenges and future opportunities. Some of the challenges include: - Dealing with ambiguity, variability and diversity of natural language in different domains, genres, styles and dialects. - Handling noisy, incomplete or inconsistent data from various sources and formats, such as web pages, social media posts or speech recordings. - Ensuring robustness, scalability and efficiency of NLP systems in real-world scenarios with large-scale data and limited resources. - Ensuring reliability, validity and fairness of NLP systems in terms of their outputs, outcomes and impacts on users and society.

What is NLG?

Natural language generation (NLG) is a software process that produces natural language output from a data set. NLG is a subfield of artificial intelligence (AI) that aims to create human-like communication from structured or unstructured data. NLG has many applications and benefits for various domains and industries, such as education, business, journalism, healthcare, and entertainment. NLG can help automate tasks that involve writing or speaking, such as generating reports, summaries, headlines, captions, scripts, emails, chatbot responses, and more. NLG can also enhance the accessibility and usability of data by transforming it into natural language that is easy to understand and interact with. How does NLG work? NLG is a complex and multi-stage process that involves several steps and techniques. The exact steps may vary depending on the specific NLG system and its goals, but a general framework can be described as follows: - Content analysis: The first step is to analyse the data source and determine what information should be included in the output. This may involve filtering, selecting, and prioritising the relevant data points based on the purpose and context of the output. - Data understanding: The next step is to interpret the data and identify patterns, trends, insights, and relationships that can be expressed in natural language. This may involve using machine learning algorithms, such as classification, clustering, regression, or sentiment analysis. - Document structuring: The third step is to create a document plan that defines the structure and organisation of the output. This may involve choosing a genre, style, tone, and format for the output, as well as deciding how to segment and order the information into paragraphs, sentences, or bullet points. - Sentence aggregation: The fourth step is to combine the information into coherent sentences that convey the main message and avoid redundancy. This may involve using linguistic techniques, such as conjunctions, anaphora, ellipsis, or paraphrasing. - Grammatical structuring: The fifth step is to apply grammatical rules and conventions to ensure that the sentences are syntactically and semantically correct. This may involve using natural language processing (NLP) tools, such as parsers, taggers, or generators. - Language presentation: The final step is to generate the output in natural language based on a template or a format that suits the target audience and medium. This may involve using natural language understanding (NLU) tools, such as lexicons, ontologies, or dialog systems. What are the challenges of NLG? NLG is a challenging task that requires a high level of linguistic knowledge and computational skills. Some of the main challenges of NLG are: - Data quality: The quality of the output depends on the quality of the input data. If the data is incomplete, inaccurate, inconsistent, or noisy, it may affect the accuracy and reliability of the output. - Data diversity: The diversity of the data sources and formats may pose difficulties for integrating and processing them. For example, different data sources may have different schemas, vocabularies, or representations that need to be aligned and harmonised. - Data complexity: The complexity of the data may make it hard to extract meaningful information and insights from it. For example, some data may be too large, too sparse, too dynamic, or too ambiguous to be easily analysed and interpreted. - Language variability: The variability of natural language may make it hard to generate natural and fluent output that matches the expectations and preferences of the users. For example, different users may have different linguistic backgrounds, dialects, registers, or styles that need to be taken into account. - Language creativity: The creativity of natural language may make it hard to generate original and engaging output that captures the attention and interest of the users. For example, some outputs may require humour, emotion, personality, or storytelling skills that are difficult to emulate by machines. What are the future trends of NLG? NLG is a rapidly evolving field that has seen significant advances in recent years thanks to the development of new technologies and methods. Some of the future trends of NLG are: - Deep learning: Deep learning is a branch of machine learning that uses neural networks to learn from large amounts of data and perform complex tasks. Deep learning has been applied to various aspects of NLG, such as data understanding, sentence aggregation, grammatical structuring and language presentation and has shown promising results in improving the quality and diversity of the output .- Multimodal NLG: Multimodal NLG is an extension of NLG that involves generating output that combines multiple modes of communication such as text image audio or video Multimedia

A Retrieval Augmented Generation (RAG) system for large language models (LLMs) is an architectural approach that enhances the performance of LLMs by incorporating external data sources as context during inference time.

RAG

Retrieval: In the RAG process, retrieval refers to the initial step where the system searches for and extracts information from a data source. This retrieval can be accomplished using a basic keyword search engine or more advanced semantic search techniques with embeddings. The primary objective is to locate the most accurate and relevant data to assist the model in answering a user’s query.

Prompt Augmentation: Following retrieval, the RAG process involves prompt augmentation. During this stage, a user’s query is combined with the retrieved documents or data to create a new prompt. This augmented prompt provides context and guides the model on how to utilize the retrieved information to address the user’s query. Often, the model receives multiple documents or data and must reason to determine which are relevant for answering the question.

Generation: The final step in the RAG process is generation. The newly created prompt, enriched with context from prompt augmentation, is introduced to the generative model. Leveraging the retrieved information, the model generates the final answer. These answers may include citations pointing to the original data sources, ensuring transparency and facilitating verification. In summary, the RAG process seamlessly combines retrieval, prompt augmentation, and generation to provide accurate and context-aware responses.

Vector Database

A vector database is a type of database designed to store and manage high-dimensional vector representations of data, known as vector embeddings or simply vectors. Here's a detailed explanation: ## What is a Vector Database? A vector database stores data in the form of vectors, which are numerical representations or embeddings of various data objects like text, images, audio, or any other type of data. These vectors are typically high-dimensional, meaning they have hundreds or thousands of dimensions, where each dimension corresponds to a specific feature or attribute of the data object. The vectors are generated by applying machine learning models, such as neural networks or word embeddings, to the raw data. The goal is to represent semantically similar data objects with vectors that are close together in the high-dimensional vector space. ## Key Characteristics - **Similarity Search**: Vector databases are optimized for similarity search, allowing users to find data objects that are similar or related to a given query vector, based on their proximity in the vector space. - **Approximate Nearest Neighbor (ANN) Algorithms**: To enable efficient similarity search, vector databases implement ANN algorithms like Hierarchical Navigable Small World (HNSW) graphs, Locality-Sensitive Hashing (LSH), or Product Quantization (PQ). - **Scalability**: Vector databases are designed to handle large volumes of high-dimensional vector data, making them suitable for applications that deal with massive amounts of unstructured or semi-structured data. - **Metadata Storage**: In addition to vectors, vector databases can store associated metadata or other structured data related to the vectors, enabling hybrid search capabilities. ## Applications Vector databases are widely used in various applications that involve similarity search, such as: - **Natural Language Processing (NLP)**: For tasks like semantic search, document clustering, and text classification. - **Computer Vision**: For image similarity search, object detection, and image classification. - **Recommendation Systems**: For finding similar products, movies, or content based on user preferences. - **Retrieval-Augmented Generation (RAG)**: Enhancing large language models by retrieving relevant context from a vector database during inference. - **Anomaly Detection**: Identifying anomalies or outliers based on their distance from normal data vectors. By leveraging vector databases, developers and data scientists can build applications that understand and operate on unstructured data in a more efficient and semantically meaningful way.