AI in Libraries: Notes for Teaching, 3rd Edition

Decoding AI Chatbots and Language Models: A Librarian's Guide

This book excerpt dives into the world of AI, focusing on chatbots, language models, and their potential impact on librarianship. Grasp the core concepts of machine learning, understand how chatbots work, and explore the future of AI in libraries.

• Learn how AI can revolutionize library services.
• Understand the capabilities and limitations of large language models.
• Discover the ethical considerations of AI in information management.
• Explore new opportunities for librarians in the age of AI.

Core Content:

1. Supervised Learning:

• Supervised learning involves training an ML model with labeled data, where a "teacher" provides the correct answers.

• Example: Optical Character Recognition (OCR): teaching a program to recognize letters by providing a training set of letters with their correct classifications.

• The program identifies patterns and correlations between input features (size, shape, color) and the correct classification (e.g., the letter ' a ').

  • Ensure the training set is adequate for the task, including a sufficient number of examples in various fonts and scripts.
  • Improve accuracy by considering the wider context of the text.

2. Unsupervised Learning:

• In unsupervised learning, the ML model is challenged to classify data into groups of similar items without any known right or wrong answers.
• Example: Clustering songs into playlists based on similarity.
• There is no training set, and the model must determine the number of clusters itself.
  • Unsupervised learning is not suitable for tasks like OCR, where accuracy is critical.

3. Semi-Supervised Learning:

• Combines supervised and unsupervised approaches.
• Useful when only a small proportion of the data is labeled, and labeling is expensive or time-consuming.

4. Self-Supervised Learning (SSL):

• Uses unsupervised learning and the data itself to label the data, then uses supervised learning on the now-labeled data.
• Relies on suitable structure or patterns in the data.
• Example: Predicting the next word in a sequence of words in a passage of English.
  • The data is not initially labeled, but scanning the text can produce a pseudo-label for the missing word.
  • It can be used with Images, determining multi-modal connections

5. Reinforcement Learning:

• Involves exploration of an environment by trial-and-error, with delayed rewards providing feedback on performance.

• The system makes a sequence of steps toward a desired goal and is rewarded or penalized for successes or failures.

• Example: A game-playing AI exploring the game, like chess, by trial and error and being rewarded for winning.

  • Commonly used in Language settings, to improve translation accuracy

6. Reinforcement Learning from Human Feedback (RLHF)

• Modern Foundation Models or Large Language Models (LLMs) often use Reinforcement Learning in a very specific way.
  • Has a base model trained by self-supervised training on most of the text. Then it has RLHF to tune the model such as answering specific questions.
  • A jury evaluates a thousand pairs of answers to determine the better and worse result to feed back into the model.

7. Chatbots:

• Dialog Agents that can provide entertainment, and are either Chit-chat agents or Task-oriented agents
  • Can potentially replace navigating web pages.
• Dialogs are complex. They can have unexpected and seemingly unpredictable sequences of events that take place between two or more participants.
  • It can be difficult to model because of the different types of agents
  • Dialogs can have different conversational settings, indexicals that can change their reference as the conversation progresses, and conversational implicature.
  • There sentences can simply be fragments to questions, that need to be looked back to earlier questions in context

Q&A

Q: What are foundation models, and why are they important?

A: Foundation models are large, pre-trained models that can be fine-tuned for various tasks. They represent a significant leap in AI capabilities because many machine learning or deep learning program can be built from them and the results will likely produce better results.

Q: What are some limitations of machine translation?

A: Machine translation can struggle with context, ambiguity, and understanding nuanced language. Resulting translations are reasonable, but not perfect.

Q: What is the difference between supervised and unsupervised learning?

A: Supervised learning uses labeled data, where a "teacher" provides the correct answers. Unsupervised learning uses unlabeled data, where the model must identify patterns and relationships without guidance.