Generative AI Glossary

• AI Model: An AI model is a complex set of parameters sometimes called “weights”. A predefined machine learning architecture may execute the model, creating a program trained to recognize patterns within a specific data set. These programs are mathematical equations that use data inputs to draw conclusions about real-world processes.
• AI Prompt: An AI prompt is any input that a user communicates to the AI to generate the intended output. A prompt can be in the form of text, a question, code snippets, or commands. With multimodal models (those that include images, videos, etc…) the prompt may contain formats other than text.The prompt is essentially the program (mostly written in natural language) that the AI model or Large Language Model interprets to produce its output.

C

• Chatbot: A chatbot is a computer program made to simulate real human conversations with humans over digital devices. While not all chatbots use AI, modern ones often function as complex digital assistants that use Natural Language Processing to understand user queries and provide automatic, personalized, and contextualized responses to them.

D

• Data Augmentation: Data augmentation is a technique commonly used in machine learning and deep learning to artificially increase the size of a dataset by applying various transformations to the existing data samples. The purpose of data augmentation is to increase the diversity of the training data, which can help improve the generalization and robustness of machine learning models. It is especially useful in scenarios where the available dataset is limited or lacks diversity.
• Deep Learning: Deep learning is a subset of machine learning that allows AI to mimic the way human brains process complex patterns and recognize objects. It uses multi-layered neural networks where each layer helps the AI better understand data patterns. Today, deep learning has various applications, with some of the most common being data refining, fraud detection, computer vision, speech processing, and natural-language understanding.

E

• Explainability/Interpretability: Interpretability refers to the extent to which someone can understand how a model works and what decision it will make. In other words, how well humans can predict any AI output for a given input. Meanwhile, explainability goes beyond this to understand how the AI made the decision. These notions are crucial for trust and transparency between humans and AI systems. This remains a vast research domain since Large Language Models were not built to explain their decisions nor cite their sources.

F

• Fine-tuning: Fine-tuning is a technique commonly used in deep learning to refine the results of a pre-trained model – typically one that has been trained on a large dataset – by adjusting its parameters to better fit the new data or task at hand. This is done to enhance and improve a model’s capabilities without training a new model from scratch. As a result, the model will be better adapted to work for specialized use cases.
• Fuzzy Logic: Fuzzy logic is a way to compute degrees of truth. Modern computers are built on Boolean logic which sees everything as a binary true or false. However with fuzzy logic, AI can identify a range of logical conclusions that resemble human reasoning. For example, in response to the question “Is it cold outside?” fuzzy logic allows for results like “very little, somewhat, moderately, fairly, very much” rather than a simple “yes or no”. This is useful for Natural Language Processing so the AI can understand semantic relations between concepts that are worded differently.without training a new model from scratch. As a result, the model will be better adapted to work for specialized use cases.

G

• Generalization: Generalization refers to a model’s ability to apply past knowledge learned from training data to new, unseen data. This determines how well the algorithm works in new settings.
• Generative Adversarial Networks (GANs): A GAN is a type of machine learning framework that consists of two neural networks: the generator and the discriminator. The generator takes random noise as input and tries to generate data samples that resemble the real data as output. The discriminator then determines if the output is actually real or fake. This process allows the generator to fine-tune its outputs and create more authentic data. The cycle continues until the discriminator is unable to identify fake data.
• Generative AI: Generative AI, also known as GenAI refers to a category of artificial intelligence that produces new content including text, images, audio or code, that mimics human creativity, making it a valuable tool for many industries. It uses datasets to study patterns and then create new, similar data in response to prompts. Often, GenAI uses Large Language Models to understand and/or produce natural language. Examples of GenAI platforms include ChatGPT or DALL-E2.
• Generative Pre-trained Transformer: Generative Pre-trained Transformer, or GPT, models are a type of Large Language Model and framework for generative AI. They use a neural network architecture and deep learning techniques to create human-like, natural language text.

H

• Human in the Loop (HITL): HITL is a process in which humans oversee and give feedback to AI models during both the training and testing phases. This is important for ensuring models produce accurate, ethical results.

K

• Keyword Search: A keyword search is based on specific words typed in a query. The search engine retrieves all documents from a database that contain one or several of the keywords in the query.

L

• Language Models (LMs): LMs are a type of AI model that predict, comprehend, generate, and interpret human language. They use Natural Language Processing and train on large data sets to decipher human languages.
• Large Language Models (LLMs): LLMs are the result of an algorithm whose training produces the model. During the execution of an LLM, it processes data and produces outputs from a specific input which may ask it to recognize, summarize, translate, predict, or generate content using very large datasets.Large language models can be adapted for use across a wide range of industries and fields. They’re most closely associated with generative AI. Developed by OpenAI, ChatGPT is one of the most recognizable Large Language Models.
• LLM Gateway: The LLM Gateway refers to the technical layer between app interfaces (chatbots, virtual assistants, in-app troubleshooting, support platforms) and the Large Language Model itself.

M

• Machine Learning (ML): Machine learning (ML) is a subtype of AI that consumes data and algorithms to build systems and enable AI to imitate how humans learn. By continuously learning, it’s goal is to improve the accuracy of its outputs.

N

• Natural Language Processing (NLP): Natural Language Processing, or NLP, is a computer program’s ability to understand spoken and written human language. NLP is used as opposed to programming languages (java, C++, Python, etc.) which are not “natural”. This allows humans to successfully interact with computers using natural sentences. NLP technology is used in Fluid Topics to enhance search.
• Neural Networks: A neural network is a type of machine learning model that replicates how the human brain functions. Its interconnected nodes or neurons are designed to process data and recognize patterns.

P

• Prompt Engineering: Prompt engineering is the process of creating and optimizing inputs for AI tools. Inputs are natural language text commands describing the task that the user wants the AI to perform. The goal is to refine inputs so that the AI completes a specific task or generates ideal outputs.
• Prompt Tuning: Prompt tuning is a cost-efficient way to improve an AI model’s outputs. While fine-tuning focuses on a model’s training data, prompt-tuning is the process of reworking the instructions given to AI so that the task is more clear. As a result, the AI produces better results without retraining the model.
• Prompt-based Learning: Prompt-based learning is a strategy within machine learning that engineers use to train Large Language Models. This is often referred to as few-shot learning. This strategy uses information from pre-trained language models so the same model can complete new tasks without needing retraining. This is useful for tasks such as text classification, machine translation, text summarization, and more.

R

• Recurrent Neural Networks (RNNs): RNNs are artificial neural networks that use sequential data inputs or time series data. These deep learning models are often used for language translation, Natural Language Processing, speech recognition, and image captioning.
• Retrieval Augmented Generation (RAG): RAG is the process of enhancing the outputs of a LLM. This is done by allowing it to retrieve data from an external knowledge base. For example, Fluid Topics’s platform enhances an LLM’s outputs with your product content. As a result, the LLM has access to specific, accurate, and up-to-date information without needing retraining.
• Role Prompting: Role prompting is a technique that prompt engineers use to control the AI output style by asking the algorithm to take on a specific role, character, or viewpoint when responding to a question or problem.

S

• Semantic Search: Semantic search is a data searching technique that aims to determine the intent and contextual meaning of the words a person is using for search. Fluid Topics leverages more than two decades of research and expertise in data enrichment and semantic search engine technology, which makes it the most relevant search engine for technical documentation on the market.

T

• Topic Modeling: Topic modeling is a technique that uses unsupervised learning to detect word clusters and phrase patterns in a document. This textual analysis is used to understand unstructured documents without the help of tags or training datasets.
• Training: Training is an iterative process where an AI model learns from a data set. The goal is to teach the AI system how to perceive and interpret the data in order to perform a specific task, make predictions, or choose decisions.
• Transformers: Transformers are essentially Recurrent Neural Networks (RNNs) with a special neuron layout unit called an attention unit. The architecture of transformers is useful for processing language because it can train faster and memorize a much longer context than typical RNNs (up to thousands of words). As a result, transformers allow you to process models with thousand of parameters, which was nearly impossible with a basic RNN architecture.

Z

• Zero-shot Prompting (aka direct prompting): Zero-shot prompting, also called direct prompting, refers to using a Large Language Model to execute a task it wasn’t trained for and without any explicit examples of the desired output. In other words, it occurs when models receive prompts that are not part of their training data, requiring them to rely on their pre-existing knowledge and natural language understanding to generate accurate results.