Glossary of AI Terminology
This is not a comprehensive glossary, but one that contains terminology that I come across as it relates to AI that is either new to me or needs clarification. Additionally, the industry is changing rapidly and new vocabulary is continuously being developed.
Adversarial Machine Learning - Adversarial Machine Learning focuses on strengthening AI models by exposing them to challenging inputs. This field encompasses techniques designed to identify and counteract attacks on AI systems. The significance of Adversarial Machine Learning lies in its role in safeguarding the security and integrity of applications powered by AI.
Agent - An Agent is a software program designed to interact with its environment, make decisions, and take actions to achieve specific objectives autonomously. It can range from simple rule-based systems to complex machine learning models. Key characteristics of an AI agent include autonomy, perception, reactivity, reasoning and decision-making, learning, communication, and being goal-oriented. AI agents operate by perceiving their environment through sensors, processing this information, making decisions based on it, and then acting using actuators. Agents can adjust to new situations, learn from experiences, and make decisions independently.
Artificial General Intelligence (AGI) - In broad terms, AGI typically means artificial intelligence that matches (or outmatches) humans on a range of tasks. But specifics about what counts as human-like, what tasks, and how many all tend to get waved away: AGI is AI, but better.
Foundation Model - An open source or commercially available LLM that can be extended with domain specific information.
Generative Pre-trained Transformer (GPT) - A GPT, or Generative Pre-trained Transformer, is an AI model primarily used for natural language processing. It's based on the Transformer architecture, which allows it to process and understand large amounts of text. The "Generative" aspect refers to its ability to generate text, and "Pre-trained" indicates that it's trained on a vast corpus of text data before being fine-tuned for specific tasks. GPT models are known for their capacity to generate coherent and contextually relevant text, answer questions, translate languages, and perform a variety of other language-related tasks.
LangChain - LangChain is an open-source collection of APIs that developers use to create LLM-powered applications. LangChain features include support for (1) Models (LLMs, chat models, and text embedding), (2) Prompts, (3) Memory Management, (4) Indexing of large datasets, (5) Chains that allow developers to integrate other tools and datasets, (6) Agents that enable decision-making, action execution, observation, and repetition until completion, and (7) Callbacks for examination and introspection.
Prompt - A Prompt is an explicit instruction given to an AI or machine learning model that it uses to produce an output. In language models a prompt is typically a text input that guides the model's subsequent generation of text. For instance, in a conversational AI, a user's question serves as a prompt to which the AI responds. In more advanced AI applications, prompts can also include images or other data forms, setting the context or task for the AI model to process and respond to accordingly. The effectiveness of an AI's response often hinges on the clarity and specificity of the prompt provided.
PyTorch - PyTorch is an open-source machine learning framework developers use to build deep learning (neural network) models. Developed by Facebook AI Research, it's known for its dynamic computational graph and user-friendly interface. Key features include GPU-accelerated tensor computing, automatic differentiation for easy backpropagation, and a comprehensive set of deep learning tools. Its flexibility and ease of use make it popular in both research and industry applications.
Reinforcement Learning - Reinforcement Learning trains AI systems to make decisions through trial and error. It uses a reward system that encourages the AI agent to learn by maximizing rewards and minimizing penalties. Reinforcement Learning is broadly applied in areas such as robotics, gaming, and autonomous vehicle development.
Reinforcement Learning from Human Feedback (RLHF) - RLHF, which includes learning from human preferences, trains a reward model using human feedback to optimize an agent's policy in machine learning. This approach improves the robustness Reinforcement Learning (RL) agents, especially in complex or unclear reward scenarios. Human feedback, typically gathered through rankings, informs the reward model, which then guides the agent's behavior. RLHF adapts to human preferences, making it suitable for tasks where quality judgment is subjective, like AI-generated storytelling. It's particularly effective in natural language processing, enhancing conversational agents, text summarization, and understanding. RLHF has been implemented in various language models like OpenAI's ChatGPT and DeepMind's Sparrow, demonstrating significant performance improvements.
Retrieval augmented generation (RAG) - RAG is a natural language processing (NLP) technique that combines the strengths of both retrieval- and generative-based artificial intelligence (AI) models. RAG AI can deliver accurate results that make the most of pre-existing knowledge but can also process and consolidate that knowledge to create unique, context-aware answers, instructions, or explanations in human-like language rather than just summarizing the retrieved data. RAG AI is different from generative AI in that it is a superset of generative AI. RAG combines the strengths of both generative AI and retrieval AI. RAG is also different from cognitive AI, which mimics the way the human brain works to get its results. RAG works by integrating retrieval-based techniques with generative-based AI models. Retrieval-based models excel at extracting information from pre-existing online sources like newspaper articles, databases, blogs, and other knowledge repositories such as Wikipedia or even internal databases. However, such models cannot produce original or unique responses. Alternatively, generative models can generate original responses that are appropriate within the context of what is being asked, but can find it difficult to maintain strict accuracy. To overcome these relative weaknesses in existing models, RAG was developed to combine their respective strengths and minimize their drawbacks. In a RAG-based AI system, a retrieval model is used to find relevant information from existing information sources while the generative model takes the retrieved information, synthesizes all the data, and shapes it into a coherent and contextually appropriate response.
Transformer Model - A transformer model is a neural network that learns context and thus meaning by tracking relationships in sequential data like the words in this sentence. Transformer models apply an evolving set of mathematical techniques, called attention or self-attention, to detect subtle ways even distant data elements in a series influence and depend on each other. Used by many current LLMs (Google, OpenAI, and others).
Transfer learning - Transfer learning is a unique technique that allows a pre-trained model to apply its knowledge to a new task. It is instrumental when you can’t curate sufficient datasets to fine-tune a model. When performing transfer learning, ML engineers freeze the model’s existing layers and append new trainable ones to the top. Reference: https://kili-technology.com/large-language-models-llms/building-domain-specific-llms-examples-and-techniques
Vectors - Vectors are arrays of numbers that represent data in a format that machine learning algorithms can process. They can encode various types of information, such as pixel intensity in images, word significance in text, or other quantifiable features of the data being modeled. By converting different forms of data into vectors, AI models can perform tasks like recognizing patterns, making predictions, and generating new data. This vectorization is a crucial step in transforming real-world, often complex data into a language that AI models can understand and learn from.
Zero-shot learning (ZSL) - Zero-shot learning (ZSL) is an approach in deep learning where a model learns to recognize and classify new types of data that it didn't see during its training. It does this by using extra information, known as auxiliary information, which helps the model make connections between what it has learned and what it hasn't seen before. For example, in a computer vision task, if a model has been trained to identify horses but not zebras, it can still recognize a zebra during testing if it knows from the auxiliary information that a zebra looks like a striped horse. This method is useful because it allows models to understand and categorize new data without needing direct examples of every possible category. ZSL is particularly important in fields like computer vision, natural language processing, and machine perception, where models often encounter new kinds of data.
Comments
Post a Comment