RAISE AI Seminar (Fall 2024)
Student Organizers: Xingyu Su, Maria Teleki, Bokun Wang
Faculty Mentors: James Caverlee, Shuiwang Ji, Tianbao Yang
Date: 9/4/24
Time: 11:30am - 12:30pm
Location: EAB B106
Title: Can LLMs Think? Investigating the Planning and Reasoning Capabilities of Large Language Models
Speakers: Sambhav Khurana & Shubham Parashar (TAMU)
Abstract: As Large Language Models (LLMs) advance in Natural Language Understanding (NLU) and Generation (NLG), the focus of research is shifting towards evaluating their capabilities beyond traditional NLP tasks. A key area of interest is their ability to perform tasks requiring planning and reasoning, which are crucial for applying LLMs in complex, real-world scenarios, including autonomous decision-making, problem-solving, and strategic thinking. Our presentation explores the critical question: Can LLMs reason? Join us to delve into this emerging field’s latest findings and future directions.
Date: 9/11/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Algorithmic and Modeling Perspectives on Distributionally Robust Reinforcement Learning
Speakers: Prof. Yan Li (TAMU ISE)
Abstract: Distributionally robust reinforcement learning refers to the problem of doing worst-case planning under potential kernel ambiguity in reinforcement learning. We present first-order methods for policy optimization and evaluation for distributionally RL, focusing on (s,a)-rectangular ambiguity sets. For policy optimization, we report 𝑂(𝑙𝑜𝑔(1/ )) iteration complexity, and 𝑂(1/ ) sample complexity for finding an 𝜖-robust policy assuming the existence of an efficient policy evaluation oracle. Then we present a construction of such an evaluation oracle and establish its 𝑂(1/ ) sample complexity for estimating the robust value function. For a class of ambiguity sets we discuss that it is possible to scale up the proposed method to large state spaces with provable convergence. We conclude by some remarks on extension of this work and, if time permitted, an important modeling perspective on rectangular ambiguity sets in distributionally robust RL.
𝜖
2
𝜖
2
𝜖
2
Date: 9/18/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Spatial Intelligence and 3D Foundation Models
Speakers: Zhiwen Fan (UT Austin)
Abstract: Future AI systems are envisioned to efficiently perceive and interact with the physical world while also being able to seamlessly re-create simulated digital environments for real-time and immersive applications. My research bridges the gap between the physical and digital worlds by enabling machines to perceive and reconstruct 3D structures from visual sensors. Moreover, my work focuses on the design of a comprehensive 3D foundation model that unifies versatile 3D problems within a single framework and can be executed differentiably. By integrating multi-modal sensory inputs and modeling temporal-spatial dynamics, we envision future intelligent agent systems that can explore and manipulate environments directly from input sensor data and learn from human actions, thereby developing spatial intelligence.
Date: 9/25/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Model Developmental Safety: A Safety-Centric Method and Applications in Vision-Language Models
Speakers: Gang Li (TAMU)
Abstract: In the real world, a learning-enabled system usually undergoes multiple cycles of model development to
enhance the system's ability to handle difficult or emerging tasks, such as ChatGPT has experienced
several cycles of development from GPT3.5 to GPT4 and GPT4o and recent GPTo1. This continual model
development process raises a significant issue that the model development for improving new tasks
may inadvertently lose capabilities of the old model. In contrast to continual learning which focuses on
trading off performance on previous tasks and new tasks, we introduce model developmental safety as
a guarantee of a learning system such that the new model should strictly preserve the existing
protected capabilities of the old model while improving its performance on target tasks. Specifically, we
present a safety-centric framework by formulating the model developmental safety as data-dependent
constraints. Under this framework, we study how to develop a pretrained vision-language model (aka
the CLIP model) for acquiring new capabilities or improving existing capabilities of image classification.
Our experiments on improving vision-based perception capabilities in autonomous driving demonstrate
the efficacy of the proposed approach.
Date: 10/02/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Geospatial Data Science for Disaster Resilience
Speakers: Prof. Lei Zou (TAMU Geography)
Abstract: In a world where natural disasters and health crises often catch us off guard, leading to unexpected catastrophic impacts, the need for a forward-thinking mindset has never been greater. Embracing the philosophy of ‘Expect the Unexpected’ is no longer merely an optional addition but an absolute imperative for facilitating resilience and well-being, especially in marginalized communities. Geospatial data science, driven by burgeoning geospatial big data (e.g., social media, mobility, and crowdsourcing), cutting-edge technologies (e.g., GeoAI, CyberGIS, and digital twins), equips us with a diverse array of tools and channels to accurately and comprehensively monitor and predict varied, unforeseen societal impacts and human needs during hazardous events. This talk will delve into the theory, methodologies, and real-world applications of harnessing geospatial big data from various sources to uncover the unforeseen, disparate societal impacts and human needs triggered by natural disasters and health crises. Additionally, this research has designed tools to effectively address disaster consequences and human needs. The developed tools and knowledge will shed light on pathways to prepare for unexpected disaster effects and empower a resilient future in changing climates and environments.
Date: 10/23/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Threshold Adaptive Sampling for Submodular Maximization under Bandit Feedback
Speakers: Wenjing Chen (TAMU)
Abstract: Efficient algorithms for discrete optimization are crucial in many real-world applications, from influence maximization in social networks to recommendation systems and large-scale facility location planning. In this talk, we will explore several applications of noisy submodular optimization and then focus on a specific challenge: submodular optimization under bandit feedback, where the noisy feedback follows an i.i.d. Sub-Gaussian distribution. I'll introduce an adaptive sampling strategy inspired by best-arm identification algorithms in multi-armed bandit theory, which can be used as a subroutine for a wide range of thresholding-based submodular optimization algorithms and offers significant improvements in sample efficiency.
Date: 10/30/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: LLMs and AI Agents for Scientific and Engineering Applications
Speakers: Zavier N. Ndum (TAMU)
Abstract: Artificial Intelligence (AI) and Machine Learning (ML) have made significant progress in nuclear science
and engineering, offering new methodologies to tackle complex problems, thereby improving predictive
capabilities, and enhancing decision-making processes. More recently, Generative AI (Large Language
Models (LLMs) and AI agents) have emerged as powerful tools for engineering applications. This
seminar introduces the fundamentals of LLMs and AI agents, emphasizing their role in Nuclear
Engineering. We will explore how AI Agents, combined with autonomous reasoning and adaptive
planning, are transforming nuclear engineering through the automation and optimization of complex
workflows that are otherwise laborious, time-consuming, and prone to human errors. A case study will
be presented for the automation of Monte Carlo (MC) Simulation workflows, with a deep dive into
Retrieval Augmented generation (RAG) acting as a virtual assistant to improve user experience,
efficiency, and productivity.
Date: 11/06/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Linear Recurrence and Linear Attention in the Era of LLMs
Speakers: Chengkai Liu (TAMU)
Abstract: As Large Language Models (LLMs) expand in scale, their computational and memory demands increase substantially. This talk introduces linear recurrence and linear attention with their applications in sequence and language modeling. By reducing complexity while maintaining accuracy, these architectures offer a path to building more resource-efficient sequence models suited for the demands in the era of LLMs.
Date: 11/13/24
Time: 11:30am - 12:30pm
Location: PETR 118
Title: Equivariant Policy Learning for Robotic Manipulation
Speakers: Dian Wang (Northeastern University)
Abstract: Despite recent advances in machine learning for robotics, current approaches often lack sample efficiency, posing a significant challenge due to the enormous time consumption to collect real-robot data. In this talk, I will present our innovative methods that tackle this challenge by leveraging the inherent symmetries in the physical environment. Specifically, I will outline a comprehensive framework of equivariant policy learning and its application across various problem settings, including reinforcement learning, behavior cloning, and grasping. Our methods not only significantly outperform state-of-the-art baselines but also achieve these results with far less data, both in simulation and in real-world scenarios. Furthermore, our approach demonstrates robustness in the presence of symmetry distortions, such as variations in camera angles.