Logo: University of Southern California

Events Calendar

Select a calendar:

Filter May Events by Event Type:




Events for May 08, 2023

  • ECE Seminar: Robust Classification under Sparse Adversarial Attacks

    ECE Seminar: Robust Classification under Sparse Adversarial Attacks

    Mon, May 08, 2023 @ 10:00 AM - 11:00 AM

    Ming Hsieh Department of Electrical and Computer Engineering

    Conferences, Lectures, & Seminars

    Speaker: Dr. Payam Delgosha, Research Assistant Professor, Computer Science Department, University of Illinois at Urbana Champaign

    Talk Title: Robust Classification under Sparse Adversarial Attacks

    Abstract: It is well-known that machine learning models are vulnerable to small but cleverly-designed adversarial perturbations that can cause misclassification. While there has been major progress in designing attacks and defenses for various adversarial settings, many fundamental and theoretical problems are yet to be resolved. In this talk, we consider classification in the presence of L0-bounded adversarial perturbations, a.k.a. sparse attacks. This setting is significantly different from other Lp-adversarial settings, with p >= 1, as the L0-ball is non-convex and highly non-smooth. In this talk, we discuss the fundamental limits of robustness in the presence of sparse attacks. In order to find an upper bound on the robust error, we introduce novel classification methods that are based on truncation. Furthermore, in order to find a lower bound on the robust error, we design a specific adversarial strategy which tries to remove the information about the true label given the adversary's budget. We discuss scenarios where the bounds match asymptotically. Motivated by the theoretical success of the proposed algorithm, we discuss how to incorporate truncation as a new component into a neural network architecture, and verify the robustness of the proposed architecture against sparse attacks through several experiments. Finally, we investigate the generalization properties and sample complexity of adversarial training in this setting.

    Biography: Payam Delgosha received his B.Sc. in Electrical Engineering and Pure Mathematics in 2012, and his M.Sc. in Electrical Engineering in 2014, both from Sharif University of Technology, Tehran, Iran. He received his Ph.D. in Electrical Engineering and Computer Sciences from the University of California at Berkeley in 2020. He joined the computer science department at the University of Illinois at Urbana Champaign as a research assistant professor in 2020. He received the 2020 IEEE Jack Keil Wolf ISIT best student paper award.

    Host: Dr. Richard M. Leahy, leahy@sipi.usc.edu

    Webcast: https://usc.zoom.us/j/97124212376?pwd=NTd0QzRzSXk3OGlzL0dIdFdXMmZYZz09

    More Information: ECE Seminar-Delgosha-050823.pdf

    WebCast Link: https://usc.zoom.us/j/97124212376?pwd=NTd0QzRzSXk3OGlzL0dIdFdXMmZYZz09

    Audiences: Everyone Is Invited

    Contact: Mayumi Thrasher

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • MoBI Seminar: Dr Bradley Voytek

    MoBI Seminar: Dr Bradley Voytek

    Mon, May 08, 2023 @ 11:00 AM - 12:00 PM

    Ming Hsieh Department of Electrical and Computer Engineering

    Conferences, Lectures, & Seminars

    Speaker: Dr Bradley Voytek, Department of Cognitive Science, Halıcıoğlu Data Science Institute, UC San Diego

    Talk Title: The physiology and function of aperiodic neural activity

    Series: MoBI Seminar Series

    Abstract: Perception, action, and cognition depend upon coordinated neural activity. This coordination operates within noisy, distributed neural networks, which themselves change with development, aging, and disease. Extensive field potential and EEG research shows that neural oscillations interact with neuronal spiking. This interaction has been proposed to be a mechanism for implementing dynamic coordination between brain regions, placing oscillations at the forefront of neuroscience research. Our work challenges our conception of what an oscillation even is. Beginning from basic theory and modeling, we show that traditional analyses conflate non-oscillatory, aperiodic activity with oscillations. To do this, we leverage neural modeling and a breadth of empirical data-”spanning human iPSC-derived cortical organoids, animal electrophysiology, invasive human EEG, and large-scale data mining. We show that, while not all things that appear oscillatory are so, the physiological information we can extract from the local field potential and EEG may nevertheless be far richer than previously thought, including nonsinusoidality of oscillation waveform shape and the aperiodic signal.

    Biography: Bradley Voytek is a Professor in the Department of Cognitive Science, the Halıcıoğlu Data Science Institute, and the Neurosciences Graduate Program at UC San Diego. He's an Alfred P. Sloan Neuroscience Research Fellow and a Kavli Fellow of the National Academies of Sciences, as well as a founding faculty member of the UC San Diego Halıcıoğlu Data Science Institute and the Undergraduate Data Science program. After his PhD at UC Berkeley, he joined Uber as their first data scientist-”when it was a 10-person startup-”where he helped build their data science strategy and team. His research lab combines large-scale data science and machine learning to study how brain regions communicate with one another, and how that communication changes with aging and disease. He is an advocate for promoting science to the public and speaks extensively with students at all grade levels about the joys of scientific research and discovery. In addition to his academic publications, his outreach work has appeared in outlets ranging from Scientific American and NPR to the San Diego Comic-Con. He is currently writing a book with neuroscientist Ashley Juavinett regarding the powerful future of data science in neuroscience discovery, though his most important contribution to science is his book with fellow neuroscientist Tim Verstynen, "Do Zombies Dream of Undead Sheep?", by Princeton University Press.

    Host: Dr Richard Leahy, leahy@sipi.usc.edu | Dr Karim Jerbi, karim.jerbi.udem@gmail.com

    Webcast: https://usc.zoom.us/j/97647013783?pwd=d1h2N3hxYUpJVU9CWlduYTZzMWNGQT09

    More Information: MoBI Seminar Flyer - 05.08.2023 Bradley Voytek.pdf

    Location: Hughes Aircraft Electrical Engineering Center (EEB) - EEB 132

    WebCast Link: https://usc.zoom.us/j/97647013783?pwd=d1h2N3hxYUpJVU9CWlduYTZzMWNGQT09

    Audiences: Everyone Is Invited

    Contact: Miki Arlen

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Mehrnoosh Mirtaheri

    Mon, May 08, 2023 @ 02:00 PM - 03:30 PM

    Thomas Lord Department of Computer Science

    University Calendar

    PhD Thesis Proposal - Mehrnoosh Mirtaheri

    Committee members: Aram Galstyan, Mohammad Rostami, Fred Morstatter, Cyrus Shahabi, Antonio Ortega

    Title: Scalable Graph-Based Models for Temporal Knowledge Graphs: Learning, Applications

    Abstract: Temporal knowledge graphs (TKGs) have emerged as a powerful tool for modeling relationships between entities in large raw text datasets. By capturing and representing these relationships in a structured, interpretable format, TKGs enable the extraction of valuable insights from vast amounts of unstructured information. Knowledge graphs allow for the identification of patterns and trends over time, enhancing our understanding of evolving connections and interactions between various entities. Moreover, they facilitate complex reasoning tasks, question answering, and data driven decision making by offering a more comprehensive view of the relationships found within the text.
    This thesis focuses on developing various models to address different challenges associated with TKG completion, such as data scarcity, scalability, and continuously evolving data. By tackling these challenges, the proposed models aim to enhance the capabilities of TKGs for analyzing and processing complex relationships within large scale text data. As a result, they enable more accurate and effective knowledge extraction and representation. The advancements presented in this thesis can greatly benefit a wide range of applications that rely on understanding the underlying structure of relationships in massive raw text datasets.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/99893841028?pwd=RlhVd29VcTltdnFCRW54dHc3ZjhrZz09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Thesis Proposal - Yufeng Yin

    Mon, May 08, 2023 @ 02:00 PM - 04:00 PM

    Thomas Lord Department of Computer Science

    University Calendar

    PhD Thesis Proposal - Yufeng Yin

    Committee Members: Mohammad Soleymani (chair), Jonathan Gratch, Mayank Kejriwal, Lynn Miller, Maja Mataric, and Xuezhe Ma

    Title: Towards Generalizable Facial Expression and Emotion Recognition

    Abstract: Facial expression and emotion recognition are critical components of human behavior understanding. However, the performance of automatic recognition methods degrades when evaluated across datasets or subjects, due to variations in humans and environmental factors. The manual coding required by supervised methods also presents significant practical limitations since they are not feasible when working with new datasets or individuals.

    In this thesis proposal, we investigate how to improve the generalization ability of the perception model through representation learning and synthetic data generation with minimal human efforts. (i) We explore unsupervised domain adaptation (UDA) approaches to obtain domain invariant and discriminative features without any target labels. The experiments show that UDA can effectively reduce the domain gap between datasets or subjects and improve model cross corpus performance for emotion recognition. (ii) We explore approaches for synthetic data generation to address the problems of the scarcity of labeled data and the diversity of subjects. Our results indicate that synthetic data can not only improve action unit (AU) detection performance but also fairness across genders, demonstrating its potential to solve AU detection in the wild. We will also discuss our future work involving unsupervised personalization on unseen speakers for emotion recognition through feature representation learning and label distribution calibration. Our proposed methods enhance model recognition accuracy and generalization ability to unseen subjects and datasets, paving the way for more effective human behavior analysis in a variety of applications.

    Audiences: Everyone Is Invited

    Contact: Melissa Ochoa

    Event Link: https://usc.zoom.us/j/94638614965?pwd=c0ozL09VVjVBNmNwRmQ4NTAybWwzdz09

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Defense - Isabel Rayas

    Mon, May 08, 2023 @ 02:00 PM - 04:00 PM

    Thomas Lord Department of Computer Science

    University Calendar

    PhD Candidate Defense: Isabel Rayas

    In-person: RTH 306

    Zoom: https://usc.zoom.us/j/95235693966?pwd=cE92UC8zejROMi8yYytyT3F5YnY1UT09

    Gaurav Sukhatme (Chair), David Caron, Stefanos Nikolaidis

    Title: Advancing Robot Autonomy for Long-Horizon Tasks

    Autonomous robots have real-world applications in diverse fields, such as mobile manipulation and environmental exploration, and many such tasks benefit from a hands-off approach in terms of human user involvement over a long task horizon. However, the level of autonomy achievable by a deployment is limited in part by the problem definition or task specification required by the system. Task specifications often require technical, low-level information that is unintuitive to describe and may result in generic solutions, burdening the user technically both before and after task completion. In this thesis, we aim to advance task specification abstraction toward the goal of increasing robot autonomy in real-world scenarios. We do so by tackling problems that address several different angles of this goal. First, we develop a way for the automatic discovery of optimal transition points between subtasks in the context of constrained mobile manipulation, removing the need for the human to hand-specify these in the task specification. We further propose a way to automatically describe constraints on robot motion by using demonstrated data as opposed to manually-defined constraints. Then, within the context of environmental exploration, we propose a flexible task specification framework, requiring just a set of quantiles of interest from the user that allows the robot to directly suggest locations in the environment for the user to study. We next systematically study the effect of including a robot team in the task specification and show that multirobot teams have the ability to improve performance under certain specification conditions, including enabling inter-robot communication. Finally, we propose methods for a communication protocol that autonomously selects useful but limited information to share with the other robots.

    Location: Ronald Tutor Hall of Engineering (RTH) - 306

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File
  • PhD Defense - Michiel De Jong

    Mon, May 08, 2023 @ 03:00 PM - 05:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    Abstract: Retrieval-augmented language models set the state-of-the-art on a broad spectrum of knowledge-intensive tasks, outperforming orders of magnitude larger models. However, such models can also be expensive for training and inference. Model performance and computational cost represent two sides of the coin: we can generally improve performance through scale at the expense of an increased computational burden. Therefore, we are really interested in pushing out the quality-compute frontier, improving performance at any given level of computational resources.

    In this dissertation, I analyze the factors that determine the computational burden of retrieval-augmented language models and propose strategies to extract a better performance-compute trade-off. The dissertation consists of three sections. The first section contains a detailed analysis of components of retrieval-augmented models and introduces methods to improve generation efficiency. The second section explores the use of dense memory to reduce the cost of encoding retrievals. Finally, the third section proposes a hybrid between dense memory and text retrieval, combining lessons from previous chapters.

    Names of the Dissertation defense committee members:
    Chair: Leana Golubchik
    Fei Sha
    Dani Yogatama
    Jacob Bien

    Venue: Zoom, https://usc.zoom.us/my/lgzoomeeting

    Location: https://usc.zoom.us/my/lgzoomeeting

    Audiences: Everyone Is Invited

    Contact: Asiroh Cham

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File