Select a calendar:
Filter September Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for September
-
PhD Dissertation Defense - Nathan Bartley
Mon, Sep 09, 2024 @ 09:00 AM - 11:00 AM
Thomas Lord Department of Computer Science
University Calendar
Title: Measuring and Mitigating Exposure Bias in Online Social Networks
Date and Time: Monday, September 9th: 9:00am - 11:00am
Location: ISI 1135
Committee Members: Kristina Lerman (Chair)< Mike Ananny, Emilio Ferrara, Fred Morstatter
Abstract: Online social platforms employ personalized feed algorithms to gather and collate messages from accounts users follow (and increasingly, from accounts they don’t follow). However, the network structure and activity of these observed accounts distorts content’s perceived popularity prior to any personalization. We call this “exposure bias:” our research focuses on quantifying it using diverse metrics, and we evaluate different algorithms that underpin personalized feeds with these metrics. We use empirical X/Twitter data and simulations in a network to assess the influence different feeds have on exposure bias. We describe a greedy algorithm based on network properties that offers insights into how to mitigate exposure bias.Location: Information Science Institute (ISI) - 1135
Audiences: Everyone Is Invited
Contact: Nathan Bartley
-
PhD Thesis Proposal - Hanchen Xie
Tue, Sep 17, 2024 @ 02:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Mitigating Environment Misalignments and Discovering Intrinsic Relations via Symbolic Alignment
Date and Time: Tuesday, September 17th, 2024: 2:00pm - 3:00pm
Location: SAL 213
Committee Members: Yue Wang (Chair), Wael AbdAlamgeed (Outisde), Aram Galstyan, Emilio Ferrara, Peter Beerel
Abstract: Modern data-driven machine learning models have achieved remarkable performance on visual recognition tasks, such as object detection and semantic segmentation. Further, when the training data is relatively comprehensive and contains intrinsic relations data, such as object dynamics relations, machine learning models can also discover such relations for downstream tasks, like dynamics prediction. However, unlike data for visual recognition, data that contains comprehensive intrinsic relations can be hard to acquire in many environments (e.g., the real world). As an alternative solution, directly deploying models trained in different environments (e.g., synthetic data) may lead to unsatisfied performance due to environment misalignment challenges. Further, finetuning any pre-trained relations discovery model is also challenging due to the absence of comprehensive data. In this thesis proposal, I will first introduce a set of datasets I designed to investigate the challenge. Then, I will discuss the environment misalignment challenges of conventional machine learning models by using the dynamics prediction tasks as a probe. Following that, I will introduce a promising research direction to mitigate such a challenge by utilizing symbolic space as the common space to align various environments so that the pre-trained relations discovery models can be directly employed and maintain satisfactory performance.
Zoom Link: https://usc.zoom.us/j/93657427715?pwd=Q2M5tyfDp7plVXp77nBU11ll3YB7Ul.1Location: Henry Salvatori Computer Science Center (SAL) - 213
Audiences: Everyone Is Invited
Contact: Hanchen Xie
-
PhD Dissertation Defense - Mianlun Zheng
Thu, Sep 19, 2024 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Real-time Multi-Resolution Neural Networks for Hand Simulation
Date and Time: Thursday, September 19th, 2024 - 2:00pm - 4:00pm
Location: EEB 539
Committee Members: Jernej Barbic, Oded Stein, and Jay Kuo
Abstract:
We present a method to simulate the hand’s musculoskeletal organs in any pose within the hand's range of motion. This method produces external hand shapes and internal organ shapes that match ground truth optical scans and medical images (MRI) in multiple scanned poses. Our system models bones, muscles, tendons, joint ligaments, and fat as separate volumetric organs that mechanically interact through contact and attachments. The match to MRI is achieved by incorporating pose-space deformation and plastic strains into the simulation. We evaluated our method by comparing it to optical scans and demonstrated that we qualitatively and quantitatively substantially decreased the error compared to previous work. Furthermore, we proposed a hard real-time machine learning shape deformer that reproduces high-quality shapes of a hero character at multiple levels of detail (LODs) and at fast speeds (a few milliseconds at most), targeting modern real-time applications.Location: Hughes Aircraft Electrical Engineering Center (EEB) - 539
Audiences: Everyone Is Invited
Contact: Mianlun Zheng
-
PhD Thesis Proposal - Jesse Zhang
Tue, Sep 24, 2024 @ 05:00 PM - 06:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Scalable Robot Adaptation with Large Pre-trained Models
Date and Time: 09/24/24 - 5:00p - 6:00p
Location: RTH 114
Committee Members: Erdem Biyik, Jesse Thomason, Joseph Lim, Daniel Seita, Somil Basil
Abstract: General robots deployed in the real world need to respond to dynamic environments and constantly learn new tasks. However, current approaches lack the ability to enable them to adapt to these ever-changing environments and tasks at scale, i.e., without extensive human supervision. My thesis proposal aims to tackle this problem by utilizing vast general knowledge stored in Large Pre-trained Models (LPTMs) to enable scalable and efficient robot adaptation. I will cover 3 fundamental paradigms in enabling robot adaptation: using LPTMs to (1) label offline data, (2) guide robots in learning new tasks online, and finally (3) adapt to new agent settings. Through extensive research in the first two paradigms and future thesis work proposed in the third, my proposal aims to produce general algorithms that will lead to robots mastering new tasks in unfamiliar environments with little human supervision.Location: Ronald Tutor Hall of Engineering (RTH) - 114
Audiences: Everyone Is Invited
Contact: Jesse Zhang
