Select a calendar:
Filter October Events by Event Type:
University Calendar
Events for October
-
PhD Thesis Proposal - Soumya Sanyal
Fri, Oct 18, 2024 @ 11:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
University Calendar
Presentation Title: Demystifying and Improving Large Language Models on Consistent Reasoning
Date and Time: 18th October, 11 AM - 12 PM
Location: OHE 114
Committee Members: Prof. Xiang Ren (Chair), Prof. Morteza Dehghani, Prof. Robin Jia, Prof. Jieyu Zhao
Presentation Abstract: Large Language Models (LLMs) have demonstrated remarkable performance on a variety of language tasks. Yet, a significant shortcoming of LLMs lies in their lack of consistency and generalization across diverse reasoning tasks. My thesis proposal aims to address this gap by systematically uncovering the limitations of LLMs in reasoning and developing methods to improve their reasoning consistency. The proposed research focuses on three core areas: (1) benchmarking the consistency of LLMs on deductive reasoning tasks, (2) accurately detecting inconsistencies in LLM reasoning across different language tasks, and (3) developing novel techniques to enhance the consistency and reliability of LLM reasoning. Through extensive research in these areas and proposed future thesis works, my proposal aims to make LLMs more consistent reasoners, ultimately minimizing the reasoning gap between humans and machines.Location: Olin Hall of Engineering (OHE) - 114
Audiences: Everyone Is Invited
Contact: Soumya Sanyal
-
PhD Dissertation Defense - Han Zhang
Mon, Oct 21, 2024 @ 09:30 AM - 11:30 AM
Thomas Lord Department of Computer Science
University Calendar
Title: Speeding up Multi-Objective Search Algorithms
Date: Oct 21, 2024
Location: SAL- Henry Salvatori Computer Science Center 213
Time: 9:30 AM - 11:30 PM
Committee members: Satish Kumar, Sven Koenig, Bistra Dilkina, Satyandra Kumar Gupta, Ariel Felner
Abstract: The multi-objective search problem is the problem of finding paths from a start state to a goal state in a graph where each edge is annotated with multiple costs. A typical task of multi-objective search is to find the Pareto frontier, that is, the set of all undominated paths from the start state to the goal state. This problem is important for many applications, such as transporting hazardous materials, where travel distance and risk are two costs that need to be considered. While researchers have developed various techniques over the past years for speeding up single-objective searches on large graphs, many of them have not been investigated in the context of multi-objective search. In this thesis, I hypothesize that one can speed up multi-objective search algorithms by applying insights gained from single-objective search algorithms after proper generalization. Specifically, I consider the following four classes of techniques that have been used to speed up single-objective search algorithms, namely, (1) by trading off solution quality with efficiency, (2) by anytime search, (3) by preprocessing techniques, and (4) by efficient data structures for time-consuming operations. We validate this hypothesis by introducing various new multi-objective search algorithms and speed-up techniques.Location: Henry Salvatori Computer Science Center (SAL) - 213
Audiences: Everyone Is Invited
Contact: Ellecia Williams
-
PhD Thesis Proposal - Weizhao Jin
Mon, Oct 21, 2024 @ 02:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Efficiency in Privacy-Preserving Computation Via Domain Knowledge
Date and Time: 10/21/24 - 2:00p - 3:00p
Location: DMC 103
Committee Members: Srivatsan Ravi, Bhaskar Krishnamachari, Harsha V. Madhyastha, Fred Morstatter
Abstract: In recent years, the importance of privacy has grown significantly due to the increasing reliance on user data for building server-side applications and services. To comply with expanding privacy regulations such as GDPR, service providers have adopted privacy-preserving primitives that maintain computational functionality while ensuring user privacy. However, a key challenge lies in integrating these privacy-preserving techniques, such as homomorphic encryption and multi-party computation, into application protocols in a way that balances the efficiency and feasibility of deployment. My thesis proposal investigates two distinct domains emphasizing privacy-preserving computation variants and proposes practical domain-knowledge-based solutions to address challenges related to overhead and protocol complexity for efficient privacy in machine learning and in networks/IoT.Location: DMC 103
Audiences: Everyone Is Invited
Contact: Felante' Charlemagne
-
PhD Dissertation Defense - Navid Hashemi
Wed, Oct 23, 2024 @ 01:30 PM - 02:50 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Scaling Control Synthesis and Verification in Autonomy Using Neurosymbolic Methods
Committee Members: Jyotirmoy Deshmukh (Chair), Bhaskar Krishnamachari, Chao Wang, Lars Lindemann, Georgios Fainekos
Date and Time: Wednesday, Oct. 23rd, 2024 - 1:30p - 2:50p
Location: DMC 111
Abstract: As the field of autonomy is embracing the use of neural networks for perception and control, Signal Temporal Logic (STL) has emerged as a popular formalism for specifying the task objectives and safety properties of such autonomous cyber-physical systems (ACPS). There are two important open problems in this research area: (1) how can we effectively train neural controllers in such ACPS applications, when the state dimensionality is higher and when the task objectives are specified over longer time horizons, and (2) how can we verify if the closed-loop system with a given neural controller satisfies given STL objectives. We review completed work in which we show how discrete-time STL (DT-STL) specifications lend themselves to a smooth neuro-symbolic encoding that enables the use of gradient-based methods for control design. We also show how a type of neuro-symbolic encoding of DT-STL specifications can be combined with neural network verification tools to provide deterministic guarantees. We also review how neural network encoding of the environment dynamics can help us combine statistical verification techniques with formal techniques for reachability analysis.Audiences: Everyone Is Invited
Contact: Navid Hashemi
-
PhD Dissertation Defense - Neal Lawton
Thu, Oct 24, 2024 @ 01:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Learning at the Local Level
Date: Thursday, October 24, 2024
Time: @ 1:00pm - 3:00pm
Location: DMC 160
Committee: Aram Galstyan, Greg Ver Steeg, Bistra Dilkina, and Assad Oberai
Abstract
In this dissertation, I present a perspective of machine learning that views feature learning as the fundamental strategy by which deep machine learning models learn to solve complex problems: when trained to perform one specific task, deep machine learning models tend to learn generalizable features that are useful for solving many different tasks. In this way, deep machine learning models learn at a local level by automatically breaking down complex problems into simple relevant subproblems. I then present a diverse collection of works that put this perspective into action to design better machine learning algorithms. These works include efficient optimization algorithms, including an algorithm for block-free parallel inference in exponential families (Chapter 2) and a novel second-order algorithm for training neural networks (Chapter 3); algorithms for efficient neural architecture search (NAS), including a morphism-based NAS algorithm for growing neural networks (Chapter 4) and a pruning-based NAS algorithm for finding more parameter-efficient PEFT architectures (Chapter 5); and algorithms for efficient fine-tuning of large language models, including an algorithm for increasing the performance of fine-tuning quantized models (Chapter 6) and a joint fine-tuning algorithm for retrieval augmented generation (RAG) pipelines (Chapter 7).Location: DMC 160
Audiences: Everyone Is Invited
Contact: Ellecia Williams
-
PhD Dissertation Defense - Yizhou Zhang
Tue, Oct 29, 2024 @ 04:30 PM - 06:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Towards Combatting Coordinated Manipulation to Online Public Opinions on Social Media
Location: KAP 138
Date and Time: October 29th, 2024: 4:30pm - 6:00pm
Committee Members: Yan Liu (Chair), Jieyu Zhao, and Kimon Drakopoulos
Abstract: Over the recent years, public opinions and online credibility have been suffering from the manipulation of campaigns that control malicious accounts to document and spread misinformation with specific narratives such as Fake News and Conspiracies. Such campaigns, also known as misinformation campaigns, are increasingly threatening various areas related to public opinions and decisions, such as politics and public health. Such threats, prominent in highly scrutinized societal events like the U.S. Presidential Elections and the COVID-19 pandemic, have significantly undermined societal trust and public interests. My thesis will discuss how to exploit machine learning to discover knowledge and skills that are helpful for combating these aforementioned social manipulation. More specifically, my thesis will present my research attempts to apply machine learning algorithms in three directions: Manipulation Source Identification, Susceptible Population Recognition and Automated Authenticity Verification. To identify the online manipulation from misinformation campaigns, my collaborators and I developed a series of neural temporal point process models that can recognize patterns of coordinated manipulators with data-driven learning and domain knowledge. To recognize users that are susceptible to specific misinformation, we developed a counterfactual neural network that can estimate the causal effect of a piece of misinformation on an individual user or a group of population. To complete our target on automated authenticity verification, we make use of the advances of Large Language Models (LLM), which can serve for generating a clarification for misinformation and reference for true information. To achieve this goal, more work on developing robust prompting engineering strategies is conducted to prevent the LLM from being deceived by the misinformation when verifying the genuineness of given text.Location: Kaprielian Hall (KAP) - 138
Audiences: Everyone Is Invited
Contact: Yizhou Zhang
-
PhD Thesis Proposal - Yi Zheng
Wed, Oct 30, 2024 @ 02:00 PM - 03:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Heuristic Search Techniques for Virtual Network Embedding
Date: October 30, 2024
Time: 2:00 pm to 3:00 pm.
Location: EEB 403
Committee Members: Satish Kumar Thittamaranahalli (chair), Sven Koenig (main advisor), Ramesh Govindan, Bhaskar Krishnamachari, Ketan Dungarshi Savla
Abstract: Virtualization is the mechanism of creating virtual representations of physical resources and is widely used in data centers and cloud computing services. It relies on the Virtual Network Embedding (VNE) problem: the cornerstone task of properly allocating the physical resources on a network to satisfy virtual requests for resources under various constraints while ensuring the quality of service. Combinatorially, the VNE problem is NP-hard to solve optimally. I hypothesize that the VNE problem can be solved efficiently and effectively in practice with the help of AI search techniques imported from the Multi-Agent Path Finding (MAPF) domain. My current work in this direction has shown that the resulting solvers can significantly outperform other state-of-the-art VNE algorithms in both solution quality and scalability. Overall, my work paves the way for using AI search techniques to address critical combinatorial problems in network resource management.Location: Hughes Aircraft Electrical Engineering Center (EEB) - 403
Audiences: Everyone Is Invited
Contact: Ellecia Williams
-
PhD Dissertation Defense - Fei Wang
Thu, Oct 31, 2024 @ 03:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title: Developing Robust and Controllable (Multimodal) Large Language Models
Date and Time: Thur, Oct 31, 2024 @ 3:00 PM - 05:00 PM
Location: RTH 211
Zoom Link: https://usc.zoom.us/my/feiwangnlp
Committee members: Aram Galstyan (Chair), Muhao Chen, Laurent Itti, Dan O’Leary
Abstract: As (multimodal) large language models (LLMs) become integral to intelligent systems, they are increasingly used in scenarios ranging from everyday applications to high-stakes domains such as healthcare, finance, and law. Consequently, there is a growing urgency to enhance the robustness and controllability of these models and mitigate critical risks in their development and deployment. This dissertation talk will introduce methods to ensure responsible outcomes from (multimodal) LLMs through three key perspectives: (1) dynamic integration of up-to-date and domain-specific knowledge, (2) robust alignment with human intents, preferences, and values, and (3) precise control over model behavior to ensure compliance with task constraints, authorization protocols, and safety requirements.Location: Ronald Tutor Hall of Engineering (RTH) - 211
Audiences: Everyone Is Invited
Contact: Ellecia Williams