[CMU AI Seminar] October 17 at 12pm (GHC 6115 & Zoom) -- Nicholas Roberts (UW Madison) -- Geometry-Aware Adaptation for Pretrained Models -- AI Seminar sponsored by SambaNova Systems

[Asher Trockman]

Dear all,

We look forward to seeing you *this Tuesday (10/17)* from *1**2:00-1:00 PM
(U.S. Eastern time)* for the next talk of this semester's *CMU AI Seminar*,
sponsored by SambaNova Systems <https://sambanova.ai/>. The seminar will be
held in GHC 6115 *with pizza provided *and will be streamed on Zoom.

To learn more about the seminar series or to see the future schedule,
please visit the seminar website <http://www.cs.cmu.edu/~aiseminar/>.

On this Tuesday (10/17), *Nicholas Roberts* (UW Madison) will be giving a
talk titled *"**Geometry-Aware Adaptation for Pretrained Models**"*.

*Title*: Geometry-Aware Adaptation for Pretrained Models

*Talk Abstract*: Machine learning models—including prominent zero-shot
models—are often trained on datasets whose labels are only a small
proportion of a larger label space. Such spaces are commonly equipped with
a metric that relates the labels via distances between them. We propose a
simple approach to exploit this information to adapt the trained model to
reliably predict new classes—or, in the case of zero-shot prediction, to
improve its performance—without any additional training. Our technique is a
drop-in replacement of the standard prediction rule, swapping arg max with
the Fréchet mean. We provide a comprehensive theoretical analysis for this
approach, studying (i) learning-theoretic results trading off label space
diameter, sample complexity, and model dimension, (ii) characterizations of
the full range of scenarios in which it is possible to predict any
unobserved class, and (iii) an optimal active learning-like next class
selection procedure to obtain optimal training classes for when it is not
possible to predict the entire range of unobserved classes. Empirically,
using easily-available external metrics, our proposed approach, LOKI, gains
up to 29.7% relative improvement over SimCLR on ImageNet and scales to
hundreds of thousands of classes. When no such metric is available, LOKI
can use self-derived metrics from class embeddings and obtains a 10.5%
improvement on pretrained zero-shot models such as CLIP.

*Speaker Bio:* Nicholas Roberts <https://nick11roberts.science> is a third
year Ph.D. student at the University of Wisconsin-Madison advised by
Frederic Sala. This past summer, he completed an internship with the
Physics of AGI research group at Microsoft Research led by Sebastien
Bubeck, working on large language models. Previously, he completed his M.S.
in the Machine Learning Department at CMU, working with Ameet Talwalkar and
Zack Lipton. Nicholas’ research is motivated by the need to democratize
machine learning and foundation models to handle the long tail of emerging
ML tasks in the sciences and beyond. In order to use these models to solve
high-impact problems in the sciences, his work aims to solve two main
challenges: (1) determine what additional data to provide them and
understand how it interacts with pretraining data, and (2) automate the
process of adapting them to new problems. To address these challenges, he
is focused on the intersection of data-centric ML (which aims to solve 1)
and automated machine learning (AutoML) (which aims to solve 2), or more
concisely data-centric AutoML. As a result of these motivating challenges,
his work on developing the foundations of data-centric AutoML has a focus
on diverse ML tasks that are far afield from standard ML domains. These
often include problems related to solving PDEs, protein folding, climate
modeling, and beyond.

*In person: *GHC 6115
*Zoom Link*:
https://cmu.zoom.us/j/99510233317?pwd=ZGx4aExNZ1FNaGY4SHI3Qlh0YjNWUT09

Thanks,
Asher Trockman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/ai-seminar-announce/attachments/20231014/e5e71253/attachment.html>