Main content start

SSP Forum: SymSys Honors Presentations 2024

Monday, June 3, 2024
Margaret Jacks Hall (Bldg. 460)
Room 126
(See description for Notes on Entry)
symsys bubbles logo

Symbolic Systems Forum
(community sessions of SYMSYS 280 - Symbolic Systems Research Seminar)

Annual Presentation of Honors Theses

Senior Honors Students
Symbolic Systems Program

Monday, June 3, 2024
12:30-2:00 pm PT
(note late ending time)
Margaret Jacks Hall (Bldg. 460), Room 126
In-person event, not recorded
(see below for entry instructions if you are not an active Stanford affiliate)

Note: Lunch is provided, if pre-ordered, only for members of SYMSYS 280, but others are welcome to bring a lunch and eat during the presentation.


12:30  Madeleine Torio Salem, "A genome-wide and brain-wide statistical investigation into the cellular and genetic architectures of mood disorders and suicidal behaviors" (Primary Advisor: Laramie Duncan, Psychiatry and Behavioral Sciences; Second Reader: Brian Knutson, Psychology)     
     Mood disorders including major depressive disorder (MDD) and bipolar disorder are leading causes of disease burden worldwide. Mood disorders are also strongly linked to risk of death by suicide. Due to challenges posed by their heterogeneous, comorbid, polygenic, and environmentally-influenced nature, the etiologies of mood disorders remain poorly understood. In recent years, human genome-wide association studies (GWAS) have linked hundreds of genetic loci to mood disorder phenotypes; the current challenge, then, is to uncover the biological mechanisms implicated by these loci. Here, we present a robust statistical approach that combines landmark GWAS and single-nucleus RNA sequencing (snRNAseq) gene expression datasets. Specifically, MAGMA v1.10 software was used to conduct gene property and conditional analysis on well-powered European-ancestry GWAS summary statistics for MDD (n=807,553; loci=102), bipolar disorder (n= 413,466; loci=64), and suicidal behaviors (n=815,178; loci=4) to identify significant independent associations with cell types as defined by the most comprehensive human brain snRNAseq dataset available to date (461 cell types, derived from 105 brain regions and 3,369,219 individual nuclei), published in 2023 by Siletti et al. Using this methodology, we mapped polygenic risk for MDD, bipolar disorder, and suicidal behavior to specific human brain cell types and their anatomical locations. Results demonstrated distinct cellular profiles for the three phenotypes that not only identified cell types and brain areas concordant with prior findings, but also identified novel cell type associations. Further analyses conducted on cell type gene expression and gene specificity data with NeuronChat and WEBGESTALT drug over-representation analysis predicted molecular interactions between implicated cell types and suggested novel drug repurposing and development opportunities. This data-driven project is “unbiased” with respect to researcher-driven hypotheses about which cell types should be associated with mood disorders, and reveals novel associations and a greater degree of cellular and molecular specificity regarding mood disorder and suicidal behavior etiology than ever previously available.

12:40  Julia Rose Fischer, "Bayesian Methods Tutorials for Psychological Research" (Primary Advisor: Nilam Ram, Communication and Psychology; Second Reader: Tobias Gerstenberg, Psychology)     
     Methods for collecting psychologically relevant data and modeling these data are rapidly advancing. Bayesian statistical methods have emerged as a set of particularly useful techniques, both within cognitive psychology and more broadly in psychological science. It is thus important for psychology researchers to have a sound understanding of these methods. In my honors thesis, I detail my development of a set of online Bayesian methods tutorials designed to be accessible to those conducting psychology research. I present a preliminary study of the educational value of one of the tutorials, as assessed by undergraduate (N = 100) and graduate (N = 5) students. The findings suggest that students generally find the tutorial useful, especially the visual elements, and report that it increases their familiarity with relevant Bayesian concepts.

12:50  Danielle Amir-Lobel, Artificial Emotions: Exploring Perceptions about the Emotional Capabilities of AI (Jonathan Levav, Business; Second Reader: Ada Aka, Business)
     Large language models such as ChatGPT are being increasingly applied in personal and professional settings to facilitate interpersonal communication. Nevertheless, people’s subjective perceptions about AI usage and its capabilities are not well understood—and these potentially flawed perceptions will determine how and in which settings AI tools are applied. Previous research indicates that, despite its objective capabilities, people may not believe that AI models can be viable tools for emotional expression, viewing emotions as unique to humans. This paper explores how beliefs about AI usage in textual production affects subjective perceptions of the resulting text. In two iterations of a novel experiment, I consistently find that people perceive a text they believe was AI-generated as performing significantly worse in affect-related attributes than the same text when they believe it was human-generated. Subsequently, these biased beliefs influence the improvements and edits people suggest. These findings highlight how people have negatively biased beliefs about AI’s emotional capabilities. This increased understanding of nuanced beliefs involved in human-AI interaction can help inform technology and policy design such that AI is leveraged appropriately and effectively.

1:00  Olivia Lee, "Leveraging Affordance Representations for Robot Learning" (Primary Advisor: Chelsea Finn, Computer Science and Electrical Engineering; Second Reader: Nick Haber, Education)
     Humans are usually capable of using prior knowledge to adapt to novel environments quickly. We can identify new instantiations of object classes and apply previously learned skills to new objects, both of which current embodied AI agents struggle with. Online reinforcement learning provides a potential solution by enabling robots to learn through trial-and-error, however current methods are sample-inefficient, lack shaping rewards, and require frequent resets. We propose a method to address the lack of shaping rewards using affordances, the action potential of objects, to create a dense shaping reward for online reinforcement learning. We leverage state-of-the-art vision-language models to predict keypoint-based affordance representations, used as intermediate dense rewards for online reinforcement learning, in addition to sparse task completion rewards. We demonstrate that dense shaping rewards speed up online reinforcement learning for robotic manipulation, facilitating success on a variety of complex object manipulation tasks.

1:10  Cole Simmons, "Babbage in Babylon: Pioneering Deep Learning Approaches to Sumerian Cuneiform" (Primary Advisor: Dan Jurafsky, Linguistic and Computer Science; Second Readers: Niek Veldhuis, Middle Eastern Languages and Cultures at UC Berkeley, and Ian Morris, Classics)
     To date, archaeologists have unearthed more than 100,000 Sumerian texts, a corpus that reflects both the apparent genesis of writing itself, as well as a rich written tradition spanning three millennia. Reading these texts, however, requires years of specialized training. I explore the application of deep learning techniques to the reading of Sumerian cuneiform, a task traditionally limited by the scarcity of high-quality datasets and the complexity of the script. I introduce three novel datasets—each designed to address different aspects of the reading process—and then use these datasets to train and evaluate a series of models. These models set state-of-the-art results in optical character recognition (OCR) and automated transliteration, highlighting the potential of leveraging modern NLP techniques in the study of ancient languages.

1:20  Thanawat Sornwamee, “Competitive option: publication game”. (Primary Advisor: Yuliy Sannikov, Business; Second Reader: Robert Wilson, Business)
     We model a multi-agent decision on the publication of a research topic as a continuous-time game, and provides numerical analysis on the shift of the system such as by changing the information disclosure policy and by changing the friction in the publication attempt.

1:30  Kunal Sinha, "Investigating Overmodification in Pragmatic Agents" (Primary Reader: Christopher Potts, Linguistics; Second Reader: Judith Degen, Linguistics)
     Speakers' tendency to include redundant information in their referential expressions appears to challenge theories of efficient and pragmatic language use. Specifically, this overmodification occurs more frequently with color modifiers than with size modifiers. While existing hypotheses seek to explain this phenomenon based on various properties of visual scenes or of human language, manipulating these properties in experiments with human participants is often infeasible or impossible. We propose to address this limitation by adopting neural networks as our learning agents. Through a series of controlled experiments, we manipulate properties of the learning environment during language acquisition to study the effect on the agents' downstream language production. We find that overmodification can emerge in pragmatic agents, but that this emergence is sensitive to various design choices regarding model architecture and hyperparameters. We discuss the effect of each of these design choices on overmodification and analyze their benefits and drawbacks, ending with a set of recommendations for future work.

1:40  Sara Bloom, "Modeling Early Childhood Attachment & Exploration Behaviors Using Reinforcement Learning" (Primary Advisor: Nick Haber, Education; Second Reader: Logan Cross, Computer Science)
     Attachment is commonly seen as a secure basis from which children can feel safe to explore the world, yet little research has been done on the relationship between attachment and exploration using computational models. This project formulates infant attachment behaviors as reward maximizing strategies learned within a minigrid environment consisting of a child agent, a parent, and toys.

1:50  Dessert refreshments


Entry to the building is open to anyone with an active Stanford ID via the card readers next to each door. If you do not have a Stanford ID, you can gain entry between 12:15 and 12:30pm ONLY by knocking on the exterior windows of room 126. These windows are to the left of the west side exterior door on the first floor of Margaret Jacks Hall, which faces the back east side of Building 420. Please do not knock on these windows after 12:30pm when the talk has started. We will not be able to come out and open the door for you at that point.