Published using Google Docs
CS 410/510 Top: Generative AI
Updated automatically every 5 minutes

CS 410/510 Top: Generative AI

Credit Hours:

4/3

Course Coordinator:

Nirupama Bulusu

Course Description:

In the modern era, the most talked about viral hit song of the week is one that appears to have been voiced by the artists Drake and The Weeknd, but without their input or consent. But how is this possible? The answer lies in Generative AI, a class of AI systems that, when trained on large data sets, can be deployed to generate text, images, videos or other outputs, all from a few well-chosen prompts.  Rapid advances in generative AI technologies, which have sprung into the popular conscious by ChatGPT, GPT-4, Bard, DALL-E and MidJourney, have the potential to revolutionize and disrupt many aspects of modern life, such as the use of these tools to design new drugs, proteins, or materials in the sciences, to provide advice to healthcare professionals in medicine, and finally, to speed up the writing of computer code, help with composing presentations, and perform summarization in the workplace. Meanwhile, governments are increasingly concerned that such systems could perpetuate biases, as well as be abused to violate citizen privacy, intellectual property laws, impersonate individuals and spread misinformation among citizens.

This course covers mathematical and computational foundations of generative modeling, as well as a review of modern tools and applications. Specific topics include variational autoencoders, generative adversarial networks, autoregressive models such as Transformers, normalizing flow models, information lattice learning, neural text decoding, prompt programming, and detection of generated content.  

Building on our understanding of Generative AI capabilities, we will study the opportunities and challenges posed by the development of this technology and its use in cognitive work. Specific applications covered will include large language models and text generation, music and creative art, engineering design, climate science, and drug discovery.

Prerequisites:

Required: Programming proficiency in Python.

Recommended: Introduction to probability theory (STAT 243, MTH 243 or equivalent)

Goals:

Upon completion of this class, students will be able to:

  1. Explain how AI algorithms can be used to generate ideas and artifacts of high quality in given domains to interpolate novel examples
  2. Explain whether AI algorithms can generate ideas and artifacts of both high quality and high novelty within specific domains that are considered as creative
  3. Understand the theoretical limits of creativity possible with generative AI
  4. Develop a comprehension of how generative AI technology could potentially impact society, spanning ethical, political, economic, and educational issues.
  5. Deploy software tools that interface with foundational models to understand how large language models can help a programmer generate, debug, and explain their source code.  
  6. Explore how to fine-tune pre-trained models for creating high performing generative AI applications.

Textbooks:

I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning, MIT Press, 2016, accompanied by several further readings and lecture notes.

References:

A selected list of references includes:

Major Topics:

Social and Ethical Issues:

Building on our understanding of Generative AI capabilities, we will study the social and ethical challenges posed by the development of this technology and its use in creative and cognitive work; using the lenses of safety, privacy, justice and explainability.