Select a calendar:
Filter April Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for April
-
PhD Defense - Eric Heiden
Fri, Apr 08, 2022 @ 11:00 AM - 01:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Eric Heiden
Time: April 8, 11am-1pm PT
Location: RTH 406 and on Zoom (https://usc.zoom.us/j/9965174023?pwd=SzlUV1NSUlZQVUNGZTNlT2h4YWpjQT09)
Committee:
Gaurav Sukhatme (chair), Jernej Barbic, S.K. Gupta, Sven Koenig, Stefanos Nikolaidis
Title: Closing the Reality Gap via Simulation-based Inference and Control
Abstract:
Simulators play a crucial role in robotics - serving as training platforms for reinforcement learning agents, informing hardware design decisions, or facilitating the prototyping of new perception and control pipelines, among many other applications. While their predictive power offers generalizability and accuracy, a core challenge lies in the mismatch between the simulated and the real world. This thesis addresses the reality gap in robotics simulators from three angles.
First, through the lens of robotic control, we investigate a robot learning pipeline that transfers skills acquired in simulation to the real world by composing task embeddings, offering a solution orthogonal to commonly used transfer learning approaches. Further, we develop an adaptive model-predictive controller that leverages a differentiable physics engine as a world representation that is updatable from sensor measurements.
Next, we develop two differentiable simulators that tackle particular problems in robotic perception and manipulation. To improve the accuracy of LiDAR sensing modules, we build a physically-based model that accounts for the measurement process in continuous-wave LiDAR sensors and the interaction of laser light with various surface materials. In robotic cutting, we introduce a differentiable simulator for the slicing of deformable objects, enabling applications in system identification and trajectory optimization.
Finally, we explore techniques that extend the capabilities of simulators to enable their construction and synchronization from real-world sensor measurements. To this end, we present a Bayesian inference algorithm that finds accurate simulation parameter distributions from trajectory-based observations. Next, we introduce a hybrid simulation approach that augments an analytical physics engine by neural networks to enable the learning of dynamical effects unaccounted for in a rigid-body simulator. In closing, we present an inference pipeline that finds the topology of articulated mechanisms from a depth or RGB video while estimating the dynamical parameters, yielding a comprehensive, interactive simulation of the real system.
WebCast Link: https://usc.zoom.us/j/9965174023?pwd=SzlUV1NSUlZQVUNGZTNlT2h4YWpjQT09
Audiences: Everyone Is Invited
Contact: Computer Science Department
-
PhD Defense - Fawad Ahmad
Wed, Apr 20, 2022 @ 08:00 AM - 10:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Fawad Ahmad
Date: April 20th, 2022
Time: 8 - 10 AM
Title: Towards Building a Live 3D Digital Twin of the World
Committee: Prof. Ramesh Govindan (chair), Prof. Konstantinos Psounis, Prof. Barath Raghavan, Prof. Muhammad Naveed
Zoom Link: https://usc.zoom.us/j/97896459075?pwd=dFJZd1N1aHUyUnlOMkFEQk56VW5zUT09
Abstract:
A live digital twin is a high-fidelity 3D representation of a physical object or scene. This digital representation continuously replicates the physical scene in near real-time. In my dissertation, I build systems that extract and leverage live digital twins of large outdoor areas. A live digital twin creates unprecedented capabilities for both computer and human consumption. It has the potential to improve safety and efficiency for autonomous driving, monitor on-going construction, and enable timely disaster relief operations etc. For humans, it means the possibility of digitally transporting to any place on the globe to live, interact and experience it in 3D like never before.
These capabilities have strict performance and accuracy requirements. Achieving these requirements is not possible today for two reasons: limited wireless bandwidths, and limited on-board compute resources. To this end, my dissertation makes two contributions as follows. First, it builds re-usable perception infrastructure to extract live 3D digital twins with low latency and high accuracy. Second, it builds end-to-end cyber-physical systems that leverage live digital twins to enable some of the capabilities mentioned above. Evaluations show that we can extract twins of large physical areas within less than one second with centimeter level accuracy. Moreover, the cyber-physical systems we build which leverage these twins enable safer and more efficient autonomous driving.WebCast Link: https://usc.zoom.us/j/97896459075?pwd=dFJZd1N1aHUyUnlOMkFEQk56VW5zUT09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Hrayr Harutyunyan
Thu, Apr 21, 2022 @ 11:00 AM - 12:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: On information captured by neural networks: connections with memorization, generalization, and learning dynamics
Abstract:
Despite their enormous capacity, modern neural networks generalize well. In this thesis proposal we use ideas from information theory to address various aspects of this phenomenon. We show that reducing label-noise information in network weights reduces memorization and improves generalization. We propose definitions for information content of data and introduce an efficient algorithm for estimating it. These definitions allow us to quantify amount of memorization of particular examples. Finally, we derive information-theoretic generalization gap bounds that depend on average information content of a single example. We demonstrate that these bounds are non-vacuous in the practical scenarios for deep learning.
Committee:
Aram Galstyan (advisor, CS)
Greg Ver Steeg (advisor, CS)
Haipeng Luo (CS)
Bistra Dilkina (CS)
Mahdi Soltanolkotabi (EE)
Location: https://usc.zoom.us/j/95264586192?pwd=WEhERDNJcGJBblFCV2RwY29IWnJhQT09
Date: Thursday, April 21, 11:00am-12:30pm.WebCast Link: https://usc.zoom.us/j/95264586192?pwd=WEhERDNJcGJBblFCV2RwY29IWnJhQT09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Matthew Fontaine
Mon, Apr 25, 2022 @ 10:30 AM - 12:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Matthew Fontaine
Committee:
Stefanos Nikolaidis (Chair, USC, Computer Science)
Bistra Dilkina (USC, Computer Science)
Gaurav Sukhatme (USC, Computer Science)
Haipeng Luo (USC, Computer Science)
Satyandra Kumar (USC, Mechanical Engineering)
Julian Togelius (NYU, Computer Science)
Title: Towards Automating the Generation of Human-Robot Interaction Scenarios
Abstract: The human robot interaction (HRI) community currently evaluates their algorithms via hand-authored user studies. When proposing a novel algorithm, each researcher designs an experimental setup to evaluate how their new algorithm performs with human subjects. While such studies are essential to evaluating how a real human will interact with a robot, robots deployed in the real world will encounter novel scenarios not evaluated in experimental settings. To discover scenarios outside of human subjects experiments, this work proposes simulating HRI scenarios, where a scenario constitutes both an environment and simulated human agents. However, both how to explore the vast space of scenarios efficiently for diverse failures and how to generate realistic scenarios that present a feasible challenge to a human-robot team are very challenging problems. This work approaches searching the continuous space of possible scenarios as a quality diversity (QD) problem, a class of optimization problem where solving algorithms find a collection of solutions spanning a space specified by measure functions, where each solution also maximizes an objective. I present methods advancing the state-of-the-art of QD algorithms, but also a new problem setting called differentiable quality diversity (DQD) that allows for the objective and measure functions to be first order differentiable. To address the realism problem, I present methods for representing scenarios via generative models that guarantee task feasibility via mixed integer linear programming. Each of these methods is combined into an efficient scenario generation framework that tests and evaluates HRI systems. Finally, the proposed work discusses techniques for increasing the complexity of the generated scenarios and for evaluating the scenarios in real-world settings with actual end users.
WebCast Link: https://usc.zoom.us/my/tehqin?pwd=Z2E1WVp3ais1Tng1V2NndTgvR1pQQT09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Chris Denniston
Wed, Apr 27, 2022 @ 03:00 PM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Chris Denniston
Title: Active Robot Perception for Understanding the Natural World
Abstract: Robot information gathering has transformed the way we take measurements of the natural world in oceans, lakes, and underground environments. This talk will highlight key areas of improvement in using measurements of the environment to inform robot autonomy. The talk will focus on improving the efficiency and usability of these systems, combining these systems with the uncertainty in the robots pose which is characteristic of these environments, and tying the end scientific goal directly to the actions the robot performs.
Time & Venue: RTH 406 and virtual (See zoom link below), 4/27 3:00PM - 4:30PM PST
Guidance Committee Members: Stefanos Nikolaidis, Heather Culbertson, Jesse Thomason, David Caron, Gaurav Sukhatme (Chair)
Location: Ronald Tutor Hall of Engineering (RTH) - 406
WebCast Link: https://usc.zoom.us/j/2869134593
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Aida Mostafazadeh Davani
Fri, Apr 29, 2022 @ 03:00 AM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Aida Mostafazadeh Davani
Time:
Friday, April 29th, 3pm, SGM 911.
Committee: Morteza Dehghani, Bistra Dilkina, Xiang Ren, and Stephen Read
Title:
Integrating Annotator Biases into Modeling Subjective Language Classification Tasks
Abstract:
Subjective annotation tasks are inherently nuanced due to annotators' individual differences in understanding of language. Training Natural Language Processing (NLP) models for making predictions in subjective tasks based on human-annotated datasets is also marked by challenges; model decisions are rarely generalizable to judgements of unseen annotators. Therefore, modeling an acceptable interpretation of subjective tasks requires integrating psychological dimensions that capture individual differences in perceiving language for each specific task. This thesis provides an alternative approach for modeling subjective NLP tasks by tailoring representations based on annotators' varying perceptions of language. First, NLP datasets for subjective tasks are investigated to demonstrate how aggregating annotation into single ground truth labels impacts the representation of different perspectives in language resources. Then, the impacts of annotators' social biases are explored to demonstrate sources for human-like biases in annotated datasets and language classifiers. And lastly, alternative approaches for incorporating annotators' individual differences into modeling their annotation behaviors are presented.Location: Seeley G. Mudd Building (SGM) - 911
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Ritesh Ahuja
Fri, Apr 29, 2022 @ 12:00 PM - 02:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Ritesh Ahuja
Dissertation Committee: Cyrus Shahabi, Bhaskar Krishnamachari, Aleksandra Korolova
Venue: Online at 12 pm -2 pm
Zoom: https://usc.zoom.us/j/7125668882
Thesis title: Differentially Private Learned Models for Location Services
Abstract:
The emergence of mobile apps (e.g., location-based services, geosocial networks, ride-sharing) led to the collection of vast amounts of location data. Publishing aggregate information about user's movements benefits research on traffic optimization, context-aware notifications and public health (e.g., disease spread). While the benefits provided by location data are indisputable, preserving location privacy is essential, since even aggregate statistics (e.g., in the form of population density maps) can leak details about individual whereabouts. To protect against privacy risks, the data curator may publish a noisy version of the dataset, transformed according to Differential Privacy (DP), the de-facto standard for releasing statistical data.
The goal of a DP mechanism is to ensure privacy while keeping the query answers as accurate as possible. Conventional approaches build DP-compliant representation of a spatial dataset by partitioning the data domain into bins, and then publishing a histogram with the noisy count of points that fall within each bin. These solutions fall short of properly capturing skewness inherent to sparse location datasets, and as a result yield poor accuracy. Instead, in this work, we propose a paradigm shift towards learned representations of data. We learn powerful machine learning (ML) models that exploit patterns within location datasets to provide more accurate location services. We focus on key location queries that are the building blocks of many processing tasks.
For population-density maps that support range count queries on snapshot releases, where each individual contributes a single location report, we design a neural database system called Spatial Neural Histograms (SNH). We model spatial data such that density features are preserved, even when DP-compliant noise is added. As such, learning can be used to also combat data modelling errors, present in DP setting. SNH employs a set of neural networks that learn from diverse regions of the dataset and at varying granularities, leading to superior accuracy. More often however, spatio-temporal density information is required for utility (e.g., in modeling COVID hotspots). As a result, the released statistics must continually capture population counts in small areas for short time periods.
When releasing multiple snapshots, individuals may contribute multiple reports to the same dataset. The ability of an adversary to breach privacy increases significantly, and a shift to user-level privacy is necessitated. We employ the pattern recognition power of neural networks, specifically Variational Auto-Encoders (VAE), to reduce the noise introduced by DP mechanisms such that accuracy is increased, while the privacy requirement is still satisfied. The system called VAE based Data Release (VDR) enables longitudinal release of location data. In addition, by limiting the number of location reports from any single user, we reduce the noise needed by DP mechanisms, while ensuring data utility is not compromised. As a post-processing step we propose statistical estimators to adjust density information to account for the fact that they are calculated on a subset of the actual data.
Lastly, recommending a user the next-location to visit is fundamentally more challenging. When considering trajectories exhibiting short and non-repetitive spatial and temporal regularity, capturing user-user correlations requires learning sophisticated ML models that have high dimensionality in the intermediate layers of the neural networks. We propose a technique called Private Location Prediction (PLP). Central to our approach is the use of the skip-gram model, and its negative sampling technique. Our work is the first to propose differentially-private learning with skip-grams. In addition, we devise data grouping techniques within the skip-gram framework that pool together trajectories from multiple users in order to accelerate learning and improve model accuracy.
Extensive experimental results on real datasets with heterogeneous characteristics show that our proposed approaches---SNH, VDR and PLP--- significantly outperform the state of the art.WebCast Link: https://usc.zoom.us/j/7125668882
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
