Select a calendar:
Filter May Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for May
-
Phd Defense - Bowen Zhang
Mon, May 02, 2022 @ 09:00 AM - 10:30 AM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Bowen Zhang
Committee chair: Prof. Leana Golubchik (CS dept.), Prof. Fei Sha,
Committee members: Prof. Laurent Itti (CS dept.), Prof. Shri Narayanan (EE dept.)
May. 2 Monday 9:00am-10:30am
Title: Visual Representation Learning with Structural Prior
Abstract: Visual representation learning is crucial for building a robust and effective visual understanding system. The goal is to build general-purpose representations to benefit multiple downstream tasks (\ie image/video classification, segmentation, retrieval, etc.) With the accessibility to large-scale datasets and the advance in complex learning methods, sophisticated neural architectures and novel training approaches have been proposed to improve visual representation. However, obtaining a versatile representation is still yet an open question. This thesis aims to leverage the visual structure to obtain more general visual representations. The key observation is that the visual components (\ie images and videos) contain structure. It can be decomposed into atomic components such as objects, attributes, clips, etc. For example, images can be decomposed into objects and can be further described by attributes. Similarly, videos can describe complex scenes composed of multiple clips or shots, where each depicts a semantically coherent event or action. As atomic components are shareable across modalities and tasks, we hope the hierarchical visual representation that is compiled from the atomic representation could achieve better generalization ability. In this thesis, we studied two scenarios to obtain the visual structures: the structure from parallel visual and text data and the pure visual domain. We achieved state-of-the-art performance on video and text retrieval, moment localization in a video corpus, image and text retrieval, action recognition, and visual storytelling with the proposed hierarchically visual representation.WebCast Link: https://usc.zoom.us/j/92058237989
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Jason Gregory
Mon, May 02, 2022 @ 01:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Thesis Proposal - Jason Gregory
Title: Decision Support Systems for Adaptive Experimental Design in Field Robotics
Abstract:
Field robots - agents that operate in complex, natural settings - have the potential for making major, tangible impact to human-robot teaming, but also face the toughest of challenges because the physical world is an unforgiving place. Experimentation plays an integral role in the research and development of fieldable systems and this must be performed in representative conditions that leverage human supervision to effectively understand capabilities of the system, assess individual components and their interactions, manage risk, and interpret results. Adaptive decision making led by a human is required for the construction of experiments, referred to as experimental design, to use insights gained from previous experiments and overcome the inherent complexity of autonomous field robotic systems and operational environments. Human experimenters, however, inherently have several shortcomings, including an inability to reason over large-scale data, sub-optimal uncertainty estimation, and biased decision making. These shortcomings can produce disastrous outcomes, including the selection of low-value experiments that introduce unnecessary delays in building system understanding as well as the selection of risky experiments that can result in major equipment damage or physical injury. To mitigate the human's drawbacks while boosting their indispensable skill sets, we seek to develop decision support systems (DSS) that can assist an experimenter during the decision making process of experiment design and reduce experimental costs by constructing more informative experiments. In this talk, I will present recent efforts in human-in-the-loop decision making for adaptive experimental design, specifically in the context of field robotics, through the development of applicable DSSs.
Committee:
Satyandra K. Gupta (advisor, Aerospace and Mechanical Engineering, Computer Science)
Gaurav Sukhatme (Computer Science)
Heather Culbertson (Computer Science)
Stefanos Nikolaidis (Computer Science)
Quan Nguyen (Aerospace and Mechanical Engineering, Computer Science)
Location: https://usc.zoom.us/j/2562545595?pwd=RnJUdWNTaStveEdQSUFiSjhkL2o5Zz09
Date: Monday, May 2, 2022 at 1-3PM.
WebCast Link: https://usc.zoom.us/j/2562545595?pwd=RnJUdWNTaStveEdQSUFiSjhkL2o5Zz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Dimitrios Stripelis
Tue, May 03, 2022 @ 04:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title
Heterogeneous Federated Learning
Abstract
Federated Learning has emerged as a standard computational paradigm for distributed training of deep learning models across data silos. The participating silos may have heterogeneous system capabilities and data specifications. In this thesis proposal, we classify these heterogeneities into computational and semantic. We present federated training policies that accelerate the convergence of the federated model by reducing the communication and processing cost required during training. We show the efficacy of these policies across a range of challenging federated environments with highly diverse data distributions. Finally, we introduce for the first time the federated data harmonization problem and present a comprehensive architecture that addresses both data harmonization, including schema mapping, data normalization, and data imputation, as well as federated learning.
Committee
José Luis Ambite (advisor, CS)
Cyrus Shahabi (co-advisor, CS)
Greg Ver Steeg (CS)
Meisam Razaviyayn (CS)
Paul Thompson (Keck)
Location
https://usc.zoom.us/j/91349269696?pwd=QzlYMFMraVc4RGpKNWxXQnhONlJpdz09
Date
Tuesday, May 3, 4:00pm-5:00pm PDT.
WebCast Link: https://usc.zoom.us/j/91349269696?pwd=QzlYMFMraVc4RGpKNWxXQnhONlJpdz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Sarah Cooney
Wed, May 04, 2022 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Sarah Cooney
Title: Toward Sustainable and Resilient Communities with HCI: Physical Structures and Socio-Cultural Factors
Committee:
Barath Raghavan (Chair), Ramesh Govindan, Bistra Dilkina, Heather Culbertson, Hajar Yazdiha (Outside Member, Sociology)
Abstract: Today more than ever we are faced with urgent, global-scale sustainability challenges. Scientists are urging everyone to contribute, and this includes the computing community. The Sustainable Human-Computer Interaction (SHCI) community has been working on these kinds of sustainability problems for almost two decades now. My research builds on the work of this community, in particular the use of Practice Theory to examine the external structures that act on individuals, often hampering their ability to make sustainable decisions. Using both qualitative and quantitative methods from human-computer interaction, my research aim is to find local solutions to global sustainability challenges while increasing community resilience and individual well-being.
First, I look at physical infrastructure through the lens of ``social infrastructure''. I build a prototype software, PatternPainter, to enable ordinary individuals to create 3D visualizations for designs of new social spaces on abandoned land in their communities. Evaluation shows this prototype allows individuals without design training to successfully create designs in 3D. I then turn to qualitative methods from HCI, specifically photo elicitation and surveys, to add context by examining how trained designers and untrained citizens view their physical environments differently in the CommYOUnity Data Study. The observations from this study can be used to inform building future technologies in the social infrastructure space. Finally, I turn to automation. I create a pipeline using the Pix2Pix style transfer algorithm and semantic segmentation to automate the process of revitalizing city streets for pedestrian use.
In parallel, I also examine religion as a socio-cultural factor impacting sustainable decision making. This builds on previous work in SHCI, which suggests that it is important to understand the social, cultural, and psychological motivations behind sustainable decision making, so that more effective technological solutions to facilitate these decisions can be built. To that end, I conducted an interview study with 14 individuals from Catholic organizations who are involved in sustainability work from a faith-based lens. I show how the insights from this study might be used to build future technology in this space.
WebCast Link: https://usc.zoom.us/j/97131163793?pwd=MXZ0NUNyd1BWK2F4U0lmUWJtNEVLdz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Victor Ardulov
Wed, May 04, 2022 @ 03:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Modeling and Regulating Human Interaction with Control Affine Dynamical Systems
Author: Victor Ardulov
Committee Members: Shrikanth Narayanan (chair), Maja MatariÄ, Thomas D Lyon
Date: May 5, 2022 3pm Pacific
Location (in-person): Ronald Tutor Hall 320
Virtual Link:
usc.zoom.us/j/98133481247?pwd=d2VqbmNhYWljcnp4d25PVEJvd3U5Zz09
Meeting ID: 981 3348 1247
Passcode: 882255
Abstract:
Human interaction is a vital component to a persons' development and well-being. These interactions enable us to over come obstacles and find resolutions that an individual might not be able to. This subject is particularly well studied in the domains of human psychology, where human behavior is diagnostically categorized and the interaction can be utilized in order to improve somebody's health.
Prior work has explored the use of computational models of human behavior to aide in the diagnostic assessment of behavioral patterns. Most recently, novel machine learning
methods and access data has invited the to study the dynamics of human interaction on a more granular time-resolution. These dynamics have been used to identify specific moments during interactions that are relevant to the over all assessment of a individuals behavior with respect to their interlocutor. By reformulating this system from the perspective of an operator that can be controlled, it invites the possibility to predict how an individual would react to a specific input from their partner, which itself lends the opportunity to plan out interventions and probes more effectively.
This dissertation presents a formulation of human interaction through a systems theoretic paradigm with a control affine element and demonstrates how these frameworks can be utilized to gain insight into improving desired outcomes and approaches towards optimizing interaction strategies. In support of the thesis, we will present the application of these techniques to the domains of forensic interviewing, psychotherapy, and neurodevelopmental diagnostics.
Bio:
Victor is a 5th year Computer Science PhD Candidate with SAIL at the University of Southern California, where his primary research is conducted in the space of modeling and guiding behavioral interactions between people. His work includes analyzing child speaking patterns to determine the truthfulness of their statements during high-stakes interviews, improving screening tools for diagnosing neuro-developmental disorders (e.g. Autism Spectrum Disorder), and building models to improve psychologists-client outcomes during therapy. Besides his work at USC, Victor has experience as a scientific advisor at Calypso AI distilling research to build AI testing software, a research engineer at Hughes Research, working on Human-AI collaborative teams, at NASA's Jet Propulsion Laboratory, working on VR tools for science and planning, and received his Bachelor's in Computer/Robotics Engineering at UC Santa Cruz, where he designed and developed assistive exoskeleton to help stroke patients.Location: 320
WebCast Link: usc.zoom.us/j/98133481247?pwd=d2VqbmNhYWljcnp4d25PVEJvd3U5Zz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Aleksei Petrenko
Thu, May 05, 2022 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
Date: 05/05/22 (Thursday)
Time: 2pm
Physical location: RTH 406 conference room
Zoom URL: https://usc.zoom.us/j/8712894950
Committee:
- Gaurav Sukhatme
- Rahul Jain
- Jesse Thomason
- Mike Zyda
- Stefanos Nikolaidis
Abstract:
We propose accelerated methods for deep reinforcement learning that enable state-of-the-art large scale experiments with simulated environments on limited hardware. We break down performance bottlenecks of reinforcement learning and discuss optimization techniques such as asynchronous experience collection for heterogeneous learning systems, large-batch rendering for high throughput simulation, and end-to-end training systems design that leverages fast GPU-based simulators. The proposal includes a case study of multiple reinforcement learning projects which heavily rely on accelerated training: from agents that learn how to execute instructions spoken in natural language to quadrotor drones trained in a simulated environment and deployed in the real world.
Location: Ronald Tutor Hall of Engineering (RTH) - 306
WebCast Link: https://usc.zoom.us/j/8712894950
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Timothy Greer
Mon, May 09, 2022 @ 01:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Creating Cross-Modal, Context-Aware Representations of Music for Downstream Tasks
PhD Candidate: Timothy Greer
Defense Committee: Prof. Shrikanth Narayanan (Chair), Prof. David Kempe,
Prof. Morteza Dehghani (External), Prof. Mohammad Soleymani, Prof. Jonas Kaplan
Date: Monday, May 9, 2022
Time: 1:00PM PST
Location: RTH 320 and Zoom
Abstract: With the ever-burgeoning market for music, film, television, and other consumable media, it has never been more important to study human music processing and multi-modal experience. Advances in computational approaches offer new ways to understand music content, and how it is experienced---both as a standalone medium and in context with other forms of media---in nuanced ways. This work identifies novel methods for representing music in a context-aware fashion which has applications in multimodal perception, music information retrieval, music emotion recognition, and recommender systems. This work also presents a music-specific transformer model that creates deep representations of audio that are enriched through a multi-task learning paradigm. By creating cross-modal, context-aware representations of music, it is possible to meaningfully capture music and media-related perception, a boon to researchers in affective computing, music information retrieval, and automatic music tagging.
Bio: Timothy is a Computer Science PhD Candidate with SAIL at the University of Southern California, where his primary research focuses on music processing and the human music listening experience. Before USC, Timothy worked at MIT Lincoln Laboratory on problems related to graph processing and natural language. In his free time, Timothy plays saxophone and keyboards in an award-winning indie-pop band called Saticöy.
Location: Ronald Tutor Hall of Engineering (RTH) - 320
WebCast Link: https://usc.zoom.us/j/92497879619?pwd=UUEyYVZsZ1FPNzFVcC9JZjNGNmZKUT09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Libby Boroson
Tue, May 10, 2022 @ 10:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Robust Loop Closures for Multi-Robot SLAM in Unstructured Environments
Date/Time: Tuesday, May 10, 2022, 10:00 am - 12:00 pm
Location: RTH 406 or on Zoom: https://usc.zoom.us/j/94676460358
Committee: Gaurav Sukhatme (Chair), Nora Ayanian, Stefanos Nikolaidis, Ketan Savla
Abstract: A key capability for a team of robots operating together in an unknown environment is building and sharing maps. As each robot explores, it must be able to build its own local map and use it for navigation. To take advantage of the benefits of working in a team, the robots should also be able to share and merge those maps. Merging these local maps into a global map requires identification of loop closures, or places where the maps overlap. However, tasks in unstructured environments, such as planetary exploration, are not well-suited to traditional visual loop closure methods like scene or object detection. These tasks may involve robots with unusual sensors, the robots may not observe the same areas, and the environment may not allow for identification of standard visual features, which all make it challenging to identify loop closures. The team may also be heterogeneous, so there may be differences in how and where the robots make their observations.
This thesis addresses the challenge of identifying robust loop closures in spite of these limitations. It includes several methods that successfully find inter-robot loop closures in challenging unstructured environments, including a method using heterogeneous sensors, a method for robots that view the world from different perspectives, and a method with ranging sensors for scenarios where robots' trajectories do not overlap. It also discusses the Autonomous PUFFER multi-robot SLAM system, a semi-real time system developed for a team of robots operating autonomously in a planetary exploration environment. Finally, it discusses how these techniques provide a framework for future multi-robot mapping in unstructured environments. The maps and systems developed will need to accurately model the environment while also supporting diverse robots and teams.Location: Ronald Tutor Hall of Engineering (RTH) - 406
WebCast Link: https://usc.zoom.us/j/94676460358
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - James Preiss
Tue, May 10, 2022 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: James Preiss
Title: Characterizing and Improving Robot Learning: A Control-Theoretic Perspective
Date/time:
May 10, 2022, 2:00-4:00pm PDT
Location:
In-person: RTH 306
Zoom: https://usc.zoom.us/j/3224457297
Committee:
Gaurav S. Sukhatme (chair)
Nora Ayanian
Ashutosh Nayyar
Stefanos Nikolaidis
Abstract:
The interface between machine learning and control has enabled robots to move outside the laboratory into challenging real-world settings. Deep reinforcement learning can scale empirically to very complex systems, but we do not yet understand precisely when and why it succeeds. Control theory focuses on simpler systems, but delivers interpretability, mathematical understanding, and guarantees. We present projects that combine these strengths.
In empirical work, we propose a framework for tasks with complex dynamics but known reward functions. We restrict the use of learning to the dynamics modeling stage, and act based on this model using traditional state-space control. We apply this framework to robotic manipulation of deformable objects.
In theoretical work, we deploy the well-understood linear quadratic regulator (LQR) problem as a test case to "look inside" algorithms and problem structure. First, we investigate how reinforcement learning algorithms depend on properties of the dynamical system by bounding the variance of the REINFORCE policy gradient estimator as a function of the LQR system matrices. Second, we introduce the framework of suboptimal covering numbers to quantify how much a good multi-system policy must change with respect to the dynamics parameters, and bound the covering number for a simple class of LQR systems.
Location: Ronald Tutor Hall of Engineering (RTH) - 306
WebCast Link: https://usc.zoom.us/j/3224457297
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Rob Brekelmans
Wed, May 18, 2022 @ 03:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Rob Brekelmans
Title: Information Geometry of Annealing Paths for Inference and Estimation
Date: Wednesday, May 18, 2022, 3:00 PM PST
Location: Seeley G Mudd (SGM) Room 226 or https://usc.zoom.us/j/91562244468
Committee: Greg Ver Steeg, Aram Galstyan, Aiichiro Nakano, Assad Oberai
Estimating normalization constants and (log) marginal likelihoods is a fundamental problem in probabilistic machine learning, playing a role in maximum likelihood learning, variational inference, and estimation of information theoretic quantities. Importance sampling, where samples from a tractable proposal distribution are reweighted based on their probability under the target density, is at the heart of many successful solutions including the evidence lower bound (ELBO), importance weighted-autoencoder (IWAE), and annealed importance sampling (AIS).
In this thesis, we provide unifying perspectives on these methods and propose methodological improvements. We propose a general approach for deriving extended state-space importance sampling bounds, leading to novel AIS and energy-based methods which can accurately estimate large values of mutual information. We consider extended state space bounds in the context of variational inference, and finally propose a new one-parameter family of annealing paths which generalize the ubiquitous geometric averaging path and can improve estimation performance on example tasks.
Location: Seeley G. Mudd Building (SGM) - 226
WebCast Link: https://usc.zoom.us/j/91562244468
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
