Select a calendar:
Filter August Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for August
-
PhD Thesis Proposal - Christopher Birmingham
Tue, Aug 23, 2022 @ 10:00 AM - 11:30 AM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Christopher Birmingham
Title: Multiparty Human-Robot Interaction: Methods For Facilitating Social Support
Date: 08/23/33 (Tuesday)
Time: 10am
Virtual Only
Zoom URL: https://usc.zoom.us/j/93201030723
Committee:
Maja Mataric
Mohammad Soleymani
Lynn Miller
Jesse Thomason
Stefanos Nikolaidis
Abstract:
Socially assistive robot (SAR) systems have the potential to improve group dynamics in social settings by facilitating social support among group members. This dissertation presents a framework for SAR facilitation of social support in the context of support groups, short, repeated group interactions in which group members provide a unique type of social support made possible by shared or similar experiences of the group members. This framework is designed to encourage group members to build trust in each other through a combination of personal disclosure and empathetic responses. Within this framework the robot facilitates social support through either role-modeling empathy and disclosure or directing and encouraging group members to give their own disclosures and empathetic responses. Facilitating social support in multiparty interactions is a novel challenge for SAR that requires sensing individuals' affect, understanding multiparty turn dynamics, and acting in a socially appropriate manner for improving social support among the group. This dissertation aims toward multiple contributions, including: novel methods for measuring social support, including computational models of empathy, disclosure, and trust; a social facilitator framework for autonomous SAR facilitation based on role modeling and directing group member social support; and turn taking methods utilizing active speaker detection and turn taking prediction for modeling multiparty turn dynamics. Together, this work aids the understanding of methods a SAR agent could use to facilitate social support in group contexts.
WebCast Link: https://usc.zoom.us/j/93201030723
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Chi Zhang
Thu, Aug 25, 2022 @ 01:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Chi Zhang
Title: Acceleration of Deep Reinforcement Learning: Efficient Algorithms and Hardware Mapping
zoom meeting ID: 243 580 6897
Committee:
Dr. Viktor Prasanna (chair)
Dr. Aiichiro Nakano (CS department)
Dr. Paul Bogdan (Non-CS department)
Abstract:
Despite the recent success of Deep Reinforcement Learning (DRL) in game playing, robotics manipulation and data center cooling, training DRL agents takes a tremendous amount of time and computation resources. This is because it requires collecting a large amount of data by interacting with the environment and a large amount of policy updates via Stochastic Gradient Descent (SGD) to converge.
To reduce the amount of data to collect, existing work adopts model-based DRL that learns a world model using the data collected by interacting with the environment. Then, it uses the world model to generate synthetic data to perform policy updates. State-of-the-art approaches generate synthetic data by uniformly sampling initial states. This generates a large amount of similar data and makes each policy update less efficient. To accelerate performing policy updates, state-of-the-art hardware mappings of DRL propose efficient customized hardware designs on FPGA. However, most of the work is only applicable for a specific range of input parameters. To further increase the speed to perform policy updates when the input batch size of the neural network is large, existing works split the input batch into multiple sub-batches and adopt multiple learners to process each sub-batch on a learner concurrently. However, the synchronization overhead including data transfer and gradient averaging significantly impairs the scalability of existing approaches.
In this work, we address these limitations by developing efficient algorithms and hardware mappings. First, we propose Maximum Entropy Model Rollouts (MEMR) that generates diverse synthetic data by prioritized sampling of the initial states such that the entropy of the generated synthetic data is maximized. We mathematically derived the maximum entropy sampling criteria assuming that the synthetic data distribution is Gaussian. To accomplish this criteria, we utilize a Prioritized Replay Buffer. Second, we propose a framework for mapping DRL algorithms with a Prioritized Replay Buffer onto heterogeneous platforms consisting of a multi-core CPU, a GPU and a FPGA. We develop specific accelerators for each primitive on CPU, FPGA and GPU. Given a DRL algorithm input parameters, our design space exploration automatically chooses the optimal mapping of various primitives based on an analytical performance model.
Finally, we propose Scalable Policy Optimization (SPO) that improves the scalability of existing multi-learner DRL by reducing the synchronization overhead via local Stochastic Gradient Descent. Our experimental evaluations on widely used benchmark environments suggest i) MEMR reduces the number of policy updates to converge compared with state-of-the-art model-based DRL; ii) our framework for hardware mapping achieves superior policy updates per second compared with other mapping methods; iii) SPO achieves nearly linear scalability as the number of learners increases.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Kuang Liu
Tue, Aug 30, 2022 @ 03:30 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Kuang Liu
Title: Simulation and Machine Learning at Exascale
Zoom ID:
994 8935 4140 (pass code: 722332)
Committee:
Aiichiro Nakano
Rajiv K. Kalia
Priya Vashishta
Abstract:
Machine learning (ML) is revolutionizing scientific research, by utilizing its ever-growing capability of knowledge discovery from massive simulation data. With the recent arrival of exascale computer, new opportunities and challenges are emerging to maximize the synergy between exaFLOPS supercomputers and ML-boosted computational science. This thesis addresses two important and intertwined problems: (i) developing efficient billion-core parallel algorithms for scientific simulations, and (ii) designing domain-specific deep neural networks for scientific data analysis. To achieve these goals, I have designed a series of algorithms with specific focus on molecular dynamics simulations and protein folding as archetypal scientific applications. Specific contributions include: (i) communication-minimizing shift-collapse algorithm for n-tuple computation on central processing units; (ii) tuple-decomposition for graphics processing unit (GPU) acceleration of n-tuple computation; (iii) graph neural network (GNN) to classify different crystalline phases; and (iv) multiscale neural network to predict protein contact map.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Binh Vu
Wed, Aug 31, 2022 @ 03:00 PM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Binh Vu
Title: Building Semantic Description of Data Sources
Committee: Craig Knoblock, Sven Koenig, Yolanda Gil, Muhao Chen, Daniel O'Leary
Abstract: A semantic description of a data source precisely describes source attributes' types and the relationships between them. Building semantic descriptions is a prerequisite to automatically publish data to knowledge graphs (KGs). Previous work on this task can be placed into two groups: learning-based and value-linked methods. The learning-based methods require manually labeled semantic descriptions to train their systems. The value-linked methods use the linked entities in a data source to discover candidate semantic descriptions by matching the values in the source with values of entities' properties; hence they are unsupervised. However, the value-linked methods need linked entities and do not work well when the source's data is not in KGs. In this thesis proposal, we propose a method to address the limitations of the value-linked methods. We hypothesize that by exploiting knowledge from web tables and KGs, we can learn semantic descriptions of data sources even when there is little overlap between the sources' data and KGs.WebCast Link: https://usc.zoom.us/j/99238519131?pwd=R2ZUYlZoYVNiNXdxTXVFU1JFZXROdz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Chung-Wei Lee
Wed, Aug 31, 2022 @ 03:00 PM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Chung-Wei Lee
Title: Online Learning and Its Applications to Games and Partially Observable Systems
Committee: Haipeng Luo (host), David Kempe, Ashutosh Nayyar, Vatsal Sharan, Jiapeng Zhang
Abstract: Online Learning is a general framework for studying sequential decision-making. I will start with its applications in solving games. In particular, Online Learning has been shown as an essential theoretical foundation when building superhuman AI in poker games. We first focus on last-iterate convergence, a favorable property for online learning algorithms in two-player zero-sum games. In normal-form games, we show optimistic multiplicative weight updates (OMWU) and optimistic gradient descent ascent (OGDA) enjoy last-iterate convergence. We then generalize the results to extensive-form games (EFGs), which model sequential actions and incomplete information that appear in card games. We show that a family of regret minimization algorithms have last-iterate convergence, with some of them based on OMWU and OGDA can even converge exponentially fast. We then consider multiplayer games, where our goal becomes minimizing the individual regret of every player. We design two algorithms achieving logarithmic regret in EFGs based on ideas including a reduction from normal-form games and usage of a self-concordant regularizer on a lifted space.
In addition to solving EFGs as an application of Online Learning to partially observable systems, we discuss other examples, including dynamic pricing and recommender systems. Specifically, we formulate the problems as bandits with graph feedback and preference elicitation and discuss our contributions therein. Finally, I will talk about future work in all directions.
WebCast Link: https://usc.zoom.us/j/96191886806?pwd=UExwOXRyaG9ETUhmaW5udEF3TjYzQT09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
