Summer Session classes begin.
Summer Session classes begin.
Thesis Defense: Catherine Chen (“Investigating Mechanisms of Relevance for Explainable Information Retrieval Systems”)
Investigating Mechanisms of Relevance for Explainable Information Retrieval Systems
As information retrieval (IR) systems, such as search engines and conversational agents, become integral to various domains, ensuring their transparency and explainability is crucial for accountability, fairness, and unbiased results. However, we still lack a clear understanding of what concepts are used to determine relevance and how those concepts are combined inside these models. This thesis argues that analyzing the underlying mechanisms of Transformer-based IR models reveals how they extract and compose relevance signals, providing the foundation for building explainable, more trustworthy, and controllable systems. First, I introduce a hypothesis-testing framework that assesses the presence and location of classical relevance notions within models, demonstrating that attention heads implementing exact-match behavior can be isolated. Second, I analyze how models extract and combine multiple relevance signals and show that they implement BM25-like behavior, identifying where relevance information is stored and pointing to opportunities for targeted controllability. Third, I apply this mechanistic approach to gender bias in retrieval, finding that gender signals overlap with term-matching signals in the model’s relevance computation, and test the potential for targeted interventions. Overall, this thesis provides a more nuanced understanding of the relevance mechanisms underlying Transformer-based IR models, enhancing explainability and creating a foundation for targeted control toward safer and more trustworthy retrieval.
Host: Professor Carsten Eickhoff
Thesis Defense: Max Merlin (“Solving the Object Scouting Problem with Locally Observable Markov Decision Processes”)
Robots acting in the real world must plan in the presence of incomplete knowledge of their environment, especially regarding task-relevant objects whose location and state may be unknown. In such cases a robot must decide which objects it may need to solve the problem, find those objects, and establish their relevant properties; I call this the object scouting problem. First, I define the Information-Seeking Macro Action (ISMA), a class of perceptual skill that uses repeated observation to reliably resolve the state of an object once it is within sensor range. Utilizing the ISMA, robots can guarantee highly accurate measurements of nearby (local) objects. Given this capability, I introduce a new class of decision processes: the Locally Observable Markov Decision Process, or LOMDP. A LOMDP allows a robot to decide which objects to observe and then to find and observe them, only ever performing task level planning using sets of fully observed objects. This enables robots to use many off-the-shelf task planners that are designed for observed environments, and which scale substantially better than planners that must account for state uncertainty. Finally I consider the problem of which objects a robot should attempt to find when solving a LOMDP. I introduce a novel task-level planning algorithm that uses the principles of goal regression and least commitment to generate partial plans that guide the robot towards finding task relevant objects. This enables us to effectively solve long horizon tasks despite object state uncertainty. Together, my work provides a practical and scalable approach to the Object Scouting problem.
Host: Professor George Konidaris
Title: Shaping Dementia Research Together: Engaging People Living With Dementia and Care Partners
Presenters:
Ellen Tambor
Karen Moss
IMPACT Community Review Panel:
Katie Brandt
Lupita Gutierrez-Parker
Freddye James
Dale Rivard
Description: Pragmatic trials have been defined as an essential approach in the national strategy to rapidly and dramatically improve dementia care. It is therefore essential that trials are aligned with what matters most to people living with dementia, their care partners and families, and the health care systems that care for them. Interventions to improve dementia care often originate from researchers rather than emerging as a direct result of asking people living with dementia and care partners to define and prioritize the problems they need addressed. This disconnect between the interventions selected by researchers for testing and people’s everyday experiences of living with dementia is one reason why some interventions do not get adopted into care delivery, despite their apparent success during trials. In this grand rounds, we present the science of engaging community partners in research on dementia care and hear directly from members of IMPACT’s community review panel.
CoPsy PhD Dissertation Defense
Speaker: Victoria Halewicz
Title: How Communities Shape What We Believe
Advisor: Steven Sloman
Juneteenth Holiday. No University exercises.