Berkeley School of Information 105th Birthday Celebration

In 1918 UC Berkeley began a full-time program in library science. Join us for this year’s celebration of the founding and history of the School of Information, School of Information Management and Systems, School of Library and Information Studies, and School of Librarianship.

More info: https://www.ischool.berkeley.edu/events/2023/105th-birthday-celebration

Program

Welcoming Remarks
Marti Hearst
Interim Dean, UC Berkeley School of Information

Achievements of Patrick Wilson
Howard D. White, Ph.D. ’74
Professor Emeritus, Drexel University

Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4
David Bamman
Associate Professor, UC Berkeley School of Information

An Interdisciplinary Framework for Evaluating Deep Facial Recognition Technologies for Forensic Applications
Justin Norman
Ph.D. Student, UC Berkeley School of Information

Presentations

Achievements of Patrick Wilson
Howard D. White, Ph.D.

This talk will feature some remarks on the contemporary bearing of Wilson’s most highly cited paper, “Situational Relevance,” published exactly half-a-century ago. It will also touch on the tradition of “foundational studies” in Berkeley’s School of Information that starts with him, and on his paradigmatic influence on the field of knowledge organization.

Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4
David Bamman, Associate Professor
UC Berkeley School of Information

As ChatGPT and other large language models are transforming the space of research and a variety of industries, understanding the data on which those models have been trained provides an important lens to understand their behavior and the risks to validity they pose to downstream tasks. In this talk, I'll describe recent research by my group in carrying out a data archaeology to infer books that are known to ChatGPT and GPT-4; we find that OpenAI models have memorized a wide collection of copyrighted materials, and that the degree of memorization is tied to the frequency with which passages of those books appear on the web. The ability of these models to memorize an unknown set of books complicates assessments of measurement validity for cultural analytics by contaminating test data; we show that models perform much better on memorized books than on non-memorized books for downstream tasks. We argue that this supports a case for open models whose training data is known.

An Interdisciplinary Framework for Evaluating Deep Facial Recognition Technologies for Forensic Applications
Justin Norman

Much has been written about flaws in facial recognition, particularly in terms of gender and racial bias. With facial recognition systems seeing widespread use in law enforcement, it is also critical that we understand its accuracy, particularly in high-stakes forensic settings.

While precision and recall can be a reasonable way to assess the overall accuracy of a recognition system, an often overlooked aspect of these measurements is the composition of the comparison group. For example, a high precision may be relatively easy to achieve if the person X has highly distinct characteristics (age, race, gender, etc.) with regard to the other people in the dataset against which they are being compared. On the other hand, the same underlying recognition system may struggle if person X shares many characteristics with the comparison group. Alternatively, most facial recognition systems make strong assumptions about the image quality, pose, obfuscations present and sizes of images presented as both source images and datasets for comparison. In the real-world there are often dramatic variations in all of these variables for any given set of images. In the classic eyewitness setting, a witness is asked to identify a suspect in a six-person lineup consisting of the suspect and five decoys with the same general characteristics and distinguishing features (facial hair, glasses, etc.) as the suspect.

We propose that a similar approach should be employed to assess the accuracy of a facial recognition system deployed in a forensic setting. This approach will ensure that the underlying facial recognition task is similar regardless of the differences or similarities between the probe and comparison faces. This allows for a more proper determination of accuracy (and thus feasibility or suitability) of the model for use in real-world, high-impact use cases.

Видео Berkeley School of Information 105th Birthday Celebration канала Berkeley School of Information