Speaker:

Asma Ghandeharioun, Senior Research Scientist, People + AI Research Team, Google DeepMind

Title:

Model Interpretability: from Illusions to Opportunities

Abstract:

With the remarkable advancements in large language models (LLMs) and concerns about issues like misinformation and harmful content, understanding their internal mechanisms is more important than ever. In this talk, I start by challenging common misconceptions about how LLMs' hidden representations influence their behavior, demonstrating several “interpretability illusions.” Next, I introduce Patchscopes, a new framework that leverages the model itself to explain its internal representations in natural language. I’ll demonstrate how Patchscopes encompasses prior interpretability methods, addresses their shortcomings, and opens up new possibilities—such as using a more capable model to explain the representations of a smaller model or correcting multihop reasoning errors. Finally, I discuss several case studies where Patchscopes helps explain the mechanics of undesirable behaviors, such as divulging latent harmful content to certain user personas or contextualization errors, and suggest mitigation strategies.

Speaker Bio:

Asma Ghandeharioun, Ph.D., is a senior research scientist with the People + AI Research team at Google DeepMind. She works on aligning AI with human values through better understanding [1] and controlling (language) models [2], uniquely by demystifying their inner workings [3] and correcting collective misconceptions along the way [4, 5]. While her current research is mostly focused on machine learning interpretability, her previous work spans conversational AI, affective computing, and, more broadly, human-centered AI. She holds a doctorate and master’s degree from MIT and a bachelor’s degree from the Sharif University of Technology. She has been trained as a computer scientist/engineer and has research experience at MIT, Google Research, Microsoft Research, Ecole Polytechnique Fédérale de Lausanne (EPFL), to name a few.

Her work has been published in premier peer-reviewed machine learning venues such as ICLR, NeurIPS, ICML, EMNLP, AAAI, ACII, and AISTATS. She has received awards at NeurIPS and her work has been featured in Wired, Wall Street Journal, and New Scientist.AV needs, anything else you may need: No particular AV or other needs.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/91475935376
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

Speaker:

Oscar Stuhler, Assistant Professor, Department of Sociology, Northwestern University

Title:

Studying Textual Representations of Social Structures

Abstract:

Textual data contain rich accounts of actions taken by different actors, of relationships, and of particular identities’ attributes. In other words, texts contain representations of social structures. Sociology has a long tradition of formally studying such structures. However, with the embrace of computational text analysis, sociologists' focus has shifted from grammar-based approaches to ones based on co-occurrence, such as word embeddings. Time to bring grammar back in! In this talk, I will first present a framework for extracting grammatically structured information like the above from text. I then show how this allows us to study larger discursive patterns by presenting both past and ongoing research. In particular, I will be talking about gender and agency in fiction writing (1850-2010) and contending conceptions of refugees in German news. I will close with some work in progress that leverages generative large language models to extend this ongoing research program.

Speaker Bio:

Oscar Stuhler is a sociologist studying discourse with formal, quantitative methods. Much of his work focuses on how to measure, analyze, and theorize textual representations of social structures. Oscar is an Assistant Professor at Northwestern University’s Department of Sociology. He completed his PhD in sociology at New York University.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/93590489365
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

Speaker:

Angeles Salles, Assistant Professor, Department of Biological Sciences, University of Illinois Chicago

Title:

Neuroethology of Bat Communication and Social Behavior

Abstract:

Bats are auditory specialists, processing acoustic signals to guide their behaviors, including prey tracking, navigation, and communication. Most bat species are very social and emit a wide array of communication calls, including food-claiming, aggressive, and appeasement calls. There is strong evidence that context plays a role in the processing of acoustic signals in humans and other animals. Yet, this process's circuits and mechanisms are still not fully understood. Bats emerge as outstanding mammalian models to explore the neural mechanisms underlying acoustic communication processing. In the lab, we work with two phylogenetically distant species of bats: Carollia perspicillata and Rousettus aegyptiacus. Though these are both frugivorous bats, these species have many differences. First, they have different echolocation mechanisms; Carollia bats are laryngeal echolocators (i.e. use their larynx to produce ultrasonic sonar signals), while Rousettus bats are lingual echolocators (i.e. use tongue clicks to produce ultrasonic sonar signals). Both species use their larynx to produce their communication signals, but their repertoire differs greatly in structure and spectral patterns. Furthermore, they have different social structures; while Carollia is a harem-forming bat with high-roost fidelity, Rousettus is promiscuous, and mating can occur with different males in a season. We are leveraging these differences to contrast and compare the underlying neural circuits for social communication and behavior across taxa. In this talk, I will provide an overview of our current work and our preliminary findings on the social behavior of these different bats and on the neural circuits used to process communication signals.

Speaker Bio:

Angie Salles is an Assistant Professor at the Department of Biological Sciences at the University of Illinois Chicago. Her career in neuroethology started during her PhD at the University of Buenos Aires, in Argentina, comparatively investigating the molecular mechanisms of learning and memory in mice and crabs. She pursued her postdoctoral work in Dr. Moss’s lab at Johns Hopkins University, investigating auditory processing in bats. Currently, Angie’s lab focuses on the neurobiology of communication and social behavior in bats.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/95924022043
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

OCTOBER MEETING: Tuesday, October 29, 2024 at 5:20pm (US Central)

LOCATION:
In person: Chambers Hall, Lower Level
600 Foster Steet, Evanston Campus

AGENDA:
TBA

SPEAKERS:
TBA

DATA SCIENCE NIGHTS are monthly talks on data science techniques or applications, organized by Northwestern University graduate students and scholars. Aspiring, beginning, and advanced data scientists are welcome! For more information: http://bit.ly/nico-dsn

Related Info

Speaker:

Xiang Cheng, Professor, Department of Chemical Engineering and Materials Science, University of Minnesota

Title:

Extraordinary Life of Microswimmers: How Does a Microorganism Swim and Navigate in Complex Environments

Abstract:

A flagellated bacterium inhabits and swims in fluids of low Reynolds number, a world, though foreign to us, is of ultimately importance to many aspects of our daily lives ranging from food production, disease prevention to environmental health. In this talk, I discuss two recent experimental works in my group on the fascinating swimming behaviors of a prominent example of flagellated bacteria, Escherichia coli. First, we study the motility of E. coli in colloidal suspensions of varying sizes and volume fractions. We find that bacteria in dilute colloidal suspensions display the quantitatively same motile behaviors as those in dilute polymer solutions, where a size-dependent motility enhancement up to 80% is observed accompanied by a strong suppression of bacterial wobbling. We then develop a simple hydrodynamic model incorporating the colloidal nature of complex fluids, which quantitatively explains bacterial wobbling dynamics and mobility enhancement in both colloidal and polymeric fluids. Second, we explore the role of multiflagellarity in maintaining the constant swimming of E. coli of different lengths. By synergizing experiments of immense sample sizes with quantitative hydrodynamic modeling and simulations, we reveal how bacteria utilize the increasing number of flagella to regulate the flagellar motor load, which leads to faster flagellar rotation neutralizing the higher fluid drag on their larger bodies. Without such a collective balancing mechanism, the swimming speed of uniflagellar bacteria generically decreases with increasing body size. Our study provides new insights into the selective advantage of multiflagellarity as a ubiquitous cellular feature of bacteria. The uncovered difference between uniflagellar and multiflagellar swimming is also important for understanding environmental influence on bacterial morphology and useful for designing artificial flagellated microswimmers.

Speaker Bio:

Xiang Cheng received his B.S. in physics from Peking University in China in 2002. He then moved to U.S. and obtained his Ph.D. in physics from the University of Chicago in 2009. He worked as a postdoctoral associate in the Department of Physics at Cornell University from 2009 to 2013. He is currently a professor at the Department of Chemical Engineering and Materials Science at the University of Minnesota. Dr. Cheng has received several academic awards, including Arthur B. Metzner Early Career Award from Society of Rheology, NSF Career Award, Packard Fellowship, DARPA Young Faculty Award, 3M non-tenured faculty award and McKnight Land-Grant Professorship. His research group studies biophysics and soft materials physics in experiments, with a special focus on the emergent flow behaviors in biological and soft matter systems. Particularly, his research interests include bacterial locomotion, hydrodynamics of active fluids, rheology of colloidal suspensions and dynamics of liquid-drop impact processes.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/95363730003
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

Speaker:

Yingdan Lu, Assistant Professor, Department of Communication Studies, Northwestern University

Title:

The Evolution of Authoritarian Propaganda in the Digital Age

Abstract:

TBA

Speaker Bio:

Yingdan Lu is an Assistant Professor in the Department of Communication Studies at Northwestern University, and co-director of the Computational Multimodal Communication Lab. Her research focuses on digital technology, political communication, and information manipulation. She uses computational and qualitative methods to understand the evolution and engagement of digital propaganda in authoritarian regimes and how individuals encounter and communicate multimodal (mis)information in AI-mediated environments. Her work has appeared in leading peer-reviewed journals across communication, political science, and human-computer interaction. Before joining Northwestern, Yingdan received her Ph.D. in Communication and a Ph.D. minor in Political Science from Stanford University.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/92346340083
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

Speaker:

Guy Aridor, Assistant Professor of Marketing, Kellogg School of Management, Northwestern University

Title:

The Value of Belief Data in Online Recommendation Systems

Abstract:

Designing algorithmic recommendation systems on online platforms is simultaneously a data collection and an algorithmic problem, though most work has focused on the algorithmic aspects. In this talk, I’ll describe several recent papers that show the value of collecting data not just on consumption behavior, but also on pre-consumption attitudes — what consumers think about items they have not consumed. I’ll discuss how an economic model of user consumption choices in environments that incorporates these attitudes can rationalize empirical consumption patterns on a movie recommendation platform, MovieLens. We then test the assumptions of and the hypotheses generated by this model in a field experiment on MovieLens that collects such belief data in order to decompose the mechanisms that drive the effectiveness of recommendations. Finally, I’ll discuss a practical procedure and a resulting open-source dataset collected via this procedure that we implement on the MovieLens platform that allows for the collection of such data at scale that can enable the incorporation of such data into the design of recommendation systems.

Speaker Bio:

Guy Aridor is an Assistant Professor of Marketing at Northwestern Kellogg School of Management and a research affiliate at CESifo. His research employs tools from economics and quantitative marketing to investigate policy and antitrust issues in the digital economy, as well as the effects of new technologies on consumer behavior. He has a particular interest in consumer privacy, recommendation systems, and social media platforms. He holds a PhD in economics from Columbia University and a BA in pure/applied mathematics, computer science, and economics from Boston University.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/99906995637
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

Speakers:

Albert Kabanda, PhD Candidate, Earth and Planetary Sciences, Northwestern University
Neelam Modi, PhD Candidate, Industrial Engineering & Management Sciences, Northwestern University
Maria Warns, PhD Candidate, Engineering Sciences & Applied Mathematics, Northwestern University

Sign Up:

Sign up to present at one of our future Lightning Talk sessions. NICO Lightning Talks are open to graduate students, postdoctoral fellows, and visiting scholars.

Location:

In person: Chambers Hall, 600 Foster Street, Lower Level
Remote option: https://northwestern.zoom.us/j/91878654083
Passcode: NICO24

About the Speaker Series:

Wednesdays@NICO is a vibrant weekly seminar series focusing broadly on the topics of complex systems, data science and network science. It brings together attendees ranging from graduate students to senior faculty who span all of the schools across Northwestern, from applied math to sociology to biology and every discipline in-between. Please visit: https://bit.ly/WedatNICO for information on future speakers.

Related Info

NOVEMBER MEETING: Tuesday, November 26, 2024 at 5:20pm (US Central)

LOCATION:
In person: Chambers Hall, Lower Level
600 Foster Steet, Evanston Campus

AGENDA:
TBA

SPEAKERS:
TBA

DATA SCIENCE NIGHTS are monthly talks on data science techniques or applications, organized by Northwestern University graduate students and scholars. Aspiring, beginning, and advanced data scientists are welcome! For more information: http://bit.ly/nico-dsn

Related Info