[Theory] New course announcement - TTIC 31270 Generative Models, Art, and Perception

[Brandie Jones via Theory]

Dear all,

This quarter, I am teaching a new graduate class at TTIC on Generative
Models, Art, and Perception that *may be great for some advanced
undergraduates*. We have several seats open still (consent required by
email to the instructor). Please feel free to forward to your students!

Many thanks,
Shiry Ginosar


TTIC 31270 - Generative Models, Art, and Perception

100 units - List B.1


This course explores the intersection of human perception and generative
models of visual data. Perception is fundamentally an inverse problem: from
raw sensory input, the brain reconstructs the latent underlying structure
of the world. Some forms of art can be seen as the corresponding forward
process: once we have inferred that structure, we create imagery that
reflects what we perceived. In doing so, our artworks often expose the
biases, constraints, and internal filters of the visual system itself.


In contrast, computational generative models operate in a forward process
without engaging in the perceptual process of solving the inverse problem
at all. They synthesize visual outputs directly by training on data,
sidestepping the inferential loop humans rely on. This class will explore
how art and computationally generated artifacts each reveal something about
the visual system’s structure, its strengths, and its limitations.


This class will examine examples of visual art that reflect perceptual
phenomena and analyze how generative models produce artifacts that,
intentionally or not, are tuned (or misaligned) to human perception. We
will cover a wide array of applications of generative models in multiple
domains of visual content creation, ranging from line drawings to movies,
and in each domain, consider the limitations of current models, primarily
when their outputs are intended to be interpreted by the human eye.


The course will cover fundamental topics in human perception, computer
vision, and generative models of visual data. On the perception side,
topics will include light and color vision, center-surround effects, edge
detectors, the “where” and “what” pathways, acuity and spatial resolution,
central and peripheral vision, depth cues (such as shape from shading and
stereopsis), and motion perception. On the computational side, we will
study the frequency domain of images, statistical models of images, and the
mechanics of modern generative models, including GANs, diffusion models,
variational autoencoders, and autoregressive approaches, as well as their
strengths, weaknesses, and use cases in visual media.


The course will consist of instructor lectures, discussions, and guest
lectures from researchers and practitioners covering specific applications
or case studies.


Prerequisites: Linear Algebra (or equivalent), Intro to Machine Learning
(or equivalent). Fundamentals of deep learning including neural networks,
backpropagation, and convolutional networks (TTIC 31230 or similar). The
hands-on component of the course will involve using Python, including
PyTorch.


Coursework: Three homework assignments and an open-ended research-based
course project. Readings will cover scientific papers and textbook chapters
from computer vision, vision science, and neuroscience.


Textbooks:

   - Torralba, Isola, Freeman, Foundations of Computer Vision
   <https://mitpress.mit.edu/9780262048972/foundations-of-computer-vision/>,
   MIT Press, 2024 (online version <https://visionbook.mit.edu/>)
   - Margaret Livingstone, Vision and Art: The Biology of Seeing, Abrams,
   2008 (online version
   <https://archive.org/details/visionartbiology0000livi_h5o9>)


Expected outcomes:

   - Develop a foundational understanding of human visual perception,
   including how the visual system processes, filters, and reconstructs the
   world around us.
   - Gain the ability to critically evaluate visual artifacts produced by
   computational models, interpreting them through the lens of perceptual
   principles.
   - Understand the fundamentals of core generative image and sequence
   models. Understand the mechanisms of these models, their strengths and
   limitations, and their practical applications.
   - Build the skills to extend existing generative models toward new
   applications, with attention to how their outputs interact with human
   perception.
   - Conduct a significant open-ended research project on a topic related
   to the course material.


*Brandie Jones *
*Executive **Administrative Assistant*
*Outreach Administrator *
Toyota Technological Institute
6045 S. Kenwood Avenue
Chicago, IL  60637
www.ttic.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/theory/attachments/20251001/f17eecdf/attachment-0001.html>