Select a calendar:
Filter June Events by Event Type:
SUNMONTUEWEDTHUFRISAT
University Calendar
Events for June
-
PhD Dissertation Defense - Chrysovalatnis Anastasiou
Wed, Jun 11, 2025 @ 12:30 PM - 02:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Recovering Trajectories From Location Data Probablistically
Date and Time: Wednesday, June 11th, 2025 | 12:30p - 2:30p
Location: PHE 106
Committee Members: Cyrus Shahabi (Chair), Jose-Luis Ambite, Marlon Boarnet
Abstract: Understanding urban mobility is crucial for effective city planning, transportation management, and the development of responsive location-based services. However, challenges associated with real-world trajectory data often significantly hamper the derivation of robust insights. These include privacy restrictions limiting access to detailed movement histories, inherent sparsity in collected data points, and uncertainty stemming from sensor inaccuracies. Existing approaches rely on deterministic assumptions (like shortest paths), or necessitate extensive calibration or large, potentially biased training datasets, hindering progress.
This thesis addresses these critical challenges by developing and evaluating a suite of novel data-driven and probabilistic methodologies. We first introduce a purely data-driven technique for time-dependent reachability analysis that leverages raw trajectory data directly, thereby bypassing the complexities of traditional graph-based. To handle data sparsity effectively, we propose time-variant, road network-constrained probabilistic models ("bridgelets"), which realistically represent the inherent uncertainty of movement between sparse location samples. Furthermore, we develop a comprehensive framework (VPE), to reliably estimate vehicle visit probabilities on road segments using observations from uncertain and potentially unreliable roadside sensors.
The practical effectiveness of the proposed methods is rigorously evaluated through extensive experiments using large-scale, real-world datasets from various cities. Quantitative and qualitative results demonstrate that our probabilistic and data-driven approaches significantly improve accuracy and efficiency compared to baseline and traditional techniques. Collectively, the contributions of this thesis provide practical, robust, and innovative tools for researchers, planners, and policymakers to gain deeper, more reliable insights into complex urban mobility dynamics, enabling more informed decision-making even when faced with prevalent data limitations.Location: Charles Lee Powell Hall (PHE) - 106
Audiences: Everyone Is Invited
Contact: Chrysovalantis Anastasiou
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Dissertation Defense - Xisen Jin
Thu, Jun 12, 2025 @ 01:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Towards Continual Learning of Language Models in the Wild
Date and Time: Thursday, June 12th, 2025 | 1:00p - 3:00p
Location: RTH 306
Committee Members: Xiang Ren (Chair), Jesse Thomason, Mahdi Soltanolkotabi
Abstract: Language language models (LLMs/LMs) have become foundations of many artificial intelligence (AI) applications, and greatly benefited users in seeking information and completing tasks. Alongside the success of LLMs, there is an increasing need to promptly update these models for new application domains, new factual knowledge, and mitigating harmful behaviors. The large-scale models and the complicated data distributions have introduced unforeseen challenges in earlier study of continual learning; at the same time, new paradigms of building models, e.g., fine-tuning open-source models, have become prevalent. These new challenges and resources create a context of continual learning of language models, which we term continual learning in the wild, that differentiates the problem from the past study.
The thesis focuses on identifying and addressing the emerging challenges in continual learning of language models. In the first part of the thesis, I propose training and evaluation protocols representative of two different goals of continual learning. I create two datasets, namely a domain-incremental research paper stream and a chronologically-ordered tweet stream, alongside downstream datasets to test model capability. In addition, I extensively evaluate existing or new continual learning algorithms for the setup and identify that knowledge distillation from past model checkpoints stands out as an effective continual learning algorithm.
In the second part of the thesis, I propose to study how merging weights of existing models can achieve the goal of fusing knowledge of multiple models without access to original training data. I propose a novel model merging algorithm, RegMean, which is simple to implement, computationally efficient, and outperforms baseline merging algorithms significantly.
In the remaining part of the thesis, I introduce my work on analyzing patterns of upstream knowledge forgetting in continual learning. I interpret significant patterns of forgetting in upstream data that arise when fine-tuning LLMs. The analysis demonstrates that accurate predictions about forgetting can be made using embedding similarity models, or matrix completion from a small set of observed occurrences of forgetting. I further illustrate how predicting forgetting can lead to the development of simple and effective continual learning algorithms.Location: Ronald Tutor Hall of Engineering (RTH) - 306
Audiences: Everyone Is Invited
Contact: Xisen Jin
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Thesis Proposal - Understanding the Role of Emotions in Online Opinion Polarization
Tue, Jun 17, 2025 @ 02:00 PM - 03:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Understanding the Role of Emotions in Online Opinion Polarization
Time and Date: Tuesday, June 17th, 2025 | 2:00p - 3:30p
Location: EEB 631
Committee Members: Kristina Lerman (Chair), Fred Morstatter, Emilio Ferrara, Lindsay Young, and Pablo Barbera
Abstract:
The thesis investigates how emotional dynamics shape ideological divisions in online discourse, particularly in the wake of issue politicization. By integrating social network analysis with cutting-edge natural language processing techniques, the study addresses three central questions: What emotions do individuals express in online discussions? How do others respond to these emotional expressions? And do these emotional exchanges influence opinion formation over time? Analyzing large-scale social media data, the thesis identifies patterns of emotional expression and reception and traces the role of affective in opinion formation. The findings illuminate the emotional undercurrents of polarization, highlighting how networked emotional exchanges can both reflect and intensify political divides in contemporary digital communication environments.Location: Hughes Aircraft Electrical Engineering Center (EEB) - 631
Audiences: Everyone Is Invited
Contact: Ashwin Rao
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Dissertation Defense - Siyi Guo
Wed, Jun 18, 2025 @ 02:00 PM - 03:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Learning Social Representations for Causal Understanding of Heterogeneous and Dynamic Online Behavor
Date and Time: Wednesday, June 8th, 2025 | 2:00pm
Location: RTH 114
Committee Members: Kristina Lerman, Emilo Ferrara, Urbashi Mitra
Abstract: Social media has become a dominant force in shaping public discourse, civic engagement, and individual behavior, offering a rich but challenging environment for studying human beliefs, behaviors, and decision-making at scale. However, modeling user behavior on these platforms is complicated by multiple challenges---the massive volume of data, multi-modality, heterogeneity across users and platforms, rapidly evolving dynamics, scarcity of annotations, and difficulty in causal analysis from observational data.
This dissertation presents a framework for monitoring, explaining, modeling, and intervening in online user behavior, that addresses the challenges of handling complex and dynamic social media data. First, to understand users' reactions to real-world events in a dynamic online environment, we propose an unsupervised methodology for detecting and explaining collective emotional reactions to events, leveraging transformer-based affect modeling and topic-guided explanations. Second, we introduce SoMeR, a self-supervised, multi-view user representation learning framework that captures diverse user behaviors across text, temporal patterns, profiles, and networks, and generalizes across platforms and tasks. Third, we develop DAMF, a domain-adaptive moral foundation inference model that enables robust supervised language modeling from heterogeneous annotated datasets through adversarial training and label distribution balancing. Finally, we propose CausalDANN, a novel framework for estimating causal effects of direct text interventions using LLM-generated counterfactuals and domain adaptation to mitigate distributional shifts.
Together, these contributions advance computational social science by addressing core challenges in tracking and modeling the temporal dynamics and heterogeneity of online user behavior. The methods developed in this thesis integrate causal inference, representation learning, and time series analysis to enable scalable, generalizable, and causally grounded understanding of social media data. In this thesis, I demonstrate their utility by applying the tools to detect and explain online reactions to offline events, identify and forecast harmful behaviors such as coordinated campaigns and hateful speech, understand the evolution of polarized discussions, infer moral values expressed in online language, and evaluate the causal impact of language on moral judgment. This work offers practical tools for researchers and policymakers seeking to better understand and engage with digital populations in complex, polarized, and fast-changing online environments.Location: Ronald Tutor Hall of Engineering (RTH) - 114
Audiences: Everyone Is Invited
Contact: Siyi Guo
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Thesis Proposal - Zhonghao Shi
Wed, Jun 18, 2025 @ 02:30 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Enabling Personalized Multimodal Interaction for Socially Assistive Robots and Agents
Date and Time: 6/18 2:30pm-4pm
Location: Ronald Tutor Hall (RTH) Room 406
Committee Members:
- Prof. Maja MatariÃÂÃÂ (Committee Chair)
- Prof. Gale Lucas
- Prof. Erdem Biyik
- Prof. David Traum
- Prof. Jonathan Tarbox (external)
Abstract:
Socially assistive robots and agents have shown great promise in providing cost-effective support across domains such as education and healthcare. However, existing socially assistive AI systems still lack the ability to engage users in real-time interactions while effectively supporting users’ cognitive learning goals and socio-emotional needs. This dissertation proposal outlines both completed and remaining work toward enabling personalized, multimodal, real-time interaction capabilities between the robots/agents and human users, particularly for users with diverse needs. The work consists of two main research directions: personalized human perception and individualized interaction design. In the area of personalized human perception, the proposal presents completed work on applying domain adaptation methods to enable personalized multimodal affect recognition for children on the autism spectrum [1], and personalized child speech recognition [2]. It also includes work on benchmarking vision-language models to better understand the performance-latency tradeoffs of frontier models for real-time human perception [3]. In the area of individualized interaction design, the proposal discusses the development of an open-source socially assistive robot learning module for accessible and personalized AI and robotics education [4], as well as work on evaluating and personalizing text-to-speech voices for delivering mental health support [5]. Finally, this proposal will present an ongoing efforts developing a personalized multimodal AI tutor to support speech therapy and learning [6].
[1] Shi, Z., Groechel, T. R., Jain, S., Chima, K., Rudovic, O., & MatariÃÂÃÂ, M. J. (2022). Toward personalized affect-aware socially assistive robot tutors for long-term interventions with children with autism. ACM Transactions on Human-Robot Interaction (THRI), 11(4), 1-28.
[2] Shi, Z., Shi, X., Feng, T., Xu, A., Narayanan, S., & MatariÃÂÃÂ, M. (2025). Examining test-time adaptation for personalized child speech recognition. Paper accepted for presentation at Interspeech 2025.
[3] Shi, Z., Zhao, E., Dennler, N., Wang, J., Xu, X., Shrestha, K., Seita, D., & MatariÃÂÃÂ, M. (2025). HRIBench: Benchmarking vision-language models for real-time human perception in human-robot interaction. Paper accepted for presentation at the International Symposium on Experimental Robotics (ISER) 2025.
[4] Shi, Z.*, O'Connell, A.*, Li, Z.*, Liu, S., Ayissi, J., Hoffman, G., ... & MatariÃÂÃÂ, M. J. (2024, March). Build your own robot friend: An open-source learning module for accessible and engaging AI education. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 21, pp. 23137-23145).
[5] Shi, Z.*, Chen, H.*, Velentza, A. M., Liu, S., Dennler, N., O'Connell, A., & Mataric, M. (2023, March). Evaluating and personalizing user-perceived quality of text-to-speech voices for delivering mindfulness meditation with different physical embodiments. In Proceedings of the 2023 ACM/IEEE International Conference on Human-Robot Interaction (pp. 516-524).
[6] Shi, Z.*, Chung, D.*, Du, Y., Zhang, J., Raina, S. & MatariÃÂÃÂ, M. J. (2025) Is AI Ready to Support Speech Therapy for Children? A Systematic Review of AI-Enabled Mobile Apps for Pediatric Speech Therapy. Accepted to Interaction Design and Children (IDC) Conference.
Location: Ronald Tutor Hall of Engineering (RTH) - 406
Audiences: Everyone Is Invited
Contact: Zhonghao Shi
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Dissertation Defense - Tingting Tang
Wed, Jun 25, 2025 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Optimizing Privacy-Preserving Machine Learning for Improved Privacy, Utility, and Efficiency Tradeoffs
Date and Time: June 25, 2025, 2 PM-4 PM
Location: EEB 539
Zoom link:https://usc.zoom.us/j/7995244109?pwd=OUp6RWhUZlFGclgyN3hkREh0Z21ldz09
Committee: Murali Annavaram (Chair), Bhaskar Krishnamachari, Sai Praneeth Karimireddy, Mengyuan Li
Abstract: Machine learning (ML) has become integral to modern data-driven systems, supporting applications from image recognition and recommendation to drug discovery and language modeling. However, the growing reliance on sensitive data during both training and inference raises critical privacy concerns. Training datasets often contain personal or proprietary information, such as social networks or medical records, while inference pipelines, especially those involving retrieval-augmented generation (RAG), may expose confidential retrieved documents or context. These risks are further exacerbated when ML models are deployed in untrusted cloud environments, making privacy-preserving machine learning (PPML) a pressing research challenge.
This dissertation investigates practical approaches to PPML, focusing on improving the tradeoff between privacy, utility, and efficiency in frameworks based on differential privacy (DP) and secure multiparty computation (MPC). For DP-based PPML, we first introduce a training algorithm for graph neural networks leveraging low-rank singular value perturbation to protect sensitive graph edges while preserving the primary graph structure. This approach achieves a significantly improved privacy-utility trade-off and demonstrates resilience to edge inference attacks. For inference-time privacy, we turn to RAG systems and propose a differentially private algorithm that extracts the most frequent keywords in the ensemble of responses using private aggregation. These keywords are then used to construct prompts to produce final responses, reducing the risk of information leakage while maintaining output quality. For MPC-based secure model inference, we present a low-rank decomposition framework that reduces the number of secure multiplications in linear layers. Two further optimizations, truncation skipping and layer concatenation, reduce overhead and improve efficiency across both 3-PC and n-PC protocols.
Together, these contributions advance the practical deployment of PPML by offering techniques that uphold formal privacy guarantees while maintaining strong model performance and system efficiency. Through a combination of low-rank approximation, semantic compression, and protocol-aware system optimizations, this dissertation offers a practical path forward for developing privacy-preserving ML systems for real-world deployment.Location: Hughes Aircraft Electrical Engineering Center (EEB) - 539
Audiences: Everyone Is Invited
Contact: Tingting Tang
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Dissertation Defense - Nan Xu
Thu, Jun 26, 2025 @ 11:00 AM - 01:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Building Trustworthy LLMs: Ensuring Inference-Time Coherence and Factuality
Date: Thursday, June 26th, 2025 | 11:00am-1:00pm
Venue: RTH 306 and Zoom https://usc.zoom.us/j/93200441032?pwd=7QqueUvIVf0WXl2LQ7AELW7ix31dNz.1
Committee Members: Xuezhe Ma (Chair), Muhao Chen, Jonathan May, Daniel E. O'Leary, Ram Nevatia
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in generating text, yet critical issues such as hallucinations and precision errors significantly impact their reliability in high-stakes tasks, warranting further in-depth research. In my thesis, I will first introduce my research aiming to tackle these challenges from the perspective of decoding, which is train-free and driven by models' own understanding of seen and generated texts. Specifically, I focus on 1) reducing undesired repetitions and off-topic generations by analyzing probability distribution of decoding steps for open-ended text generation and 2) mitigating hallucinations by studying models' uncertainty against user prompts for false-premise question answering.
Motivated by the in-context learning (ICL) capabilities of Large Language Models (LLMs), multimodal LLMs incorporating an additional visual modality have demonstrated similar ICL abilities when provided with multiple image-text pairs as demonstrations. As the final part of this thesis, I will investigate whether multimodal LLMs can reliably perform a broad range of tasks without additional fine-tuning, including tasks that were not encountered during pretraining or that may even conflict with the pretraining data.
Location: Ronald Tutor Hall of Engineering (RTH) - 306
Audiences: Everyone Is Invited
Contact: Nan Xu
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.
