SSP Forum: Senior Honors Students

The
Symbolic Systems Forum
presents

Annual Presentation of Honors Projects
featuring
Senior Honors Students
Symbolic Systems Program

Monday, June 1, 2020
12:30-1:40 pm [NOTE late ending time]
Join from PC, Mac, Linux, iOS or Android: https://stanford.zoom.us/j/92146755757
    Meeting ID: 921 4675 5757
    SIP: 92146755757 [at] zoomcrc.com (92146755757[at]zoomcrc[dot]com)

SCHEDULE:

12:30 Sophie Regan, "Semantic Adaptation in Quantifier Use in Preschool Aged Children" (Advisor: Mike Frank, Psychology; Second Reader: Judith Degen, Linguistics)
Abstract: How flexible are children’s semantic representations? It is unknown whether they understand that others may have semantic boundaries that are different from their own. Adults update their expectations about how a speaker uses quantifiers after exposure to the speaker (Yildirim et al., 2016). Here, we explore whether this ability is present in preschool-aged children. In Experiment 1, we show that preschoolers have adult-like expectations about how a generic speaker would use the quantifiers some (< 50%) and many (>50%). In Experiment 2, forty 4 and 5-year-olds (mean = 4.6) were exposed to a speaker who was biased to either prefer using some or many in a situation with a proportion of 50%. After exposure, participants updated their expectations about the use of some and many, such that they aligned better with the exposure speaker’s usage, suggesting that preschoolers are able to engage in semantic adaptation. 

12:40 Tania Dhaliwal, "Characterizing the Development of Relational Reasoning in India" (Advisor: Mike Frank, Psychology; Second Reader: George Kachergis, Psychology)
Abstract: The development of relational reasoning has been shown to follow different learning curves. English speakers in the U.S. show a puzzling decline in relational ability between 18 months and 4 years old, whereas Mandarin speakers in China show steady improvement over this period. However, limited progress has been made in surveying the development of relational reasoning across more varied cultural contexts. In order to understand the mechanisms through which variation in the learning environment influences the development of relational reasoning, we examine early relational reasoning in Punjabi speakers in India since they share cultural and linguistic features with both China and the U.S. In experiment 1, we find that although 3-year-olds in India – like 3-year-olds in the U.S. – perform at chance on a relational-match-to-sample task, they exhibit an intermediate performance without any significant difference from their peers in either China or the U.S. In experiments 2a and 2b, we examine the baseline differences in bias toward relational or object-based solutions and the cultural factors that might be affecting the development of relational reasoning, and find conflicting results. Preliminary analyses on 3-4-year-olds in India and the U.S. suggest that while children in both cultures exhibit a bias towards the object-based solutions in an ambiguous match-to-sample task, Indian children are relatively more relation-focused than their U.S. peers in the culture tasks. Together, these results highlight the complexity in the development of relational reasoning and call for further research to pull apart the factors that account for the differences seen in different contexts.

12:50 Jñani Crawford, "Validation and Generalization of Pixel-wise Relevance in CNNs Trained for Face Classification" (Advisor: Kalanit Grill-Spector, Psychology; Second Reader: Sonia Poltoratski, Psychology)
Abstract: The increased use of convolutional neural networks for face recognition in science, governance, and broader society has led to an increased need for methods that can codify how these 'black box' decisions are made. For a given high-performing neural network, such a method should explain which parts of the model's learned classification strategy are robust to random initializations or spurious correlations in input data. In this paper, the decompositional pixel-wise attribution method of layer-wise relevance propagation (LRP) is applied to this end using several classes of VGG-16 models that vary in their pretraining (ImageNet or VGGFace) and finetuning task (gender or identity classification). Following this, the models are evaluated on relevance-based occluded images to analyze to what degree each model's relevance distributions vary with and generalize across the type of pretraining dataset, the finetuning task, and random initializations. We find that relevance distributions produced by these various types of VGG-16 models prove generally stable across random initializations, and can generalize across finetuning tasks. However, there is much less generalization across types of pretraining data. These results suggest that differently pretrained networks have both shared and exclusive relevant pixels as found using LRP, and that finetuning on the same task also increases the overlap of these relevant pixels.

1:00 Ellie Bowen, "Charlotte Salomon and the Message in a Bottle: A Philosophical Meditation on Charlotte Salomon’s Leben? oder Theater? and the Meaning of Artistic Creation in the Face of Mortality, Told in Three Parts" (Advisor: Alexander Nemerov, Art and Art History; Second Reader: Robert P. Harrison, French and Italian)
Abstract: This thesis project explores the art of Charlotte Salomon, who created a gesamtkunstwerk entitled “Life? or Theatre?” — chronicling her life story — in the midst of WWII, while hiding in exile from the Nazis. Applying the philosophies of Nietzsche and Kierkegaard, this thesis argues that Salomon’s project was an artistic leap of faith, evidencing her belief in life despite the darkness that surrounded her. It explores the redemptive aspects of art amidst great loss, and how this artistic redemption differs from notions of Judeo-Christian redemption. Finally, the thesis talks about what it means to be a receiver of a work of art, analogizing the experience to that of the serendipity of receiving a message in a bottle from the sea. 

1:10 Nathan Lee, "Universal Basic Participation: Reforming Campaign Finance by Expanding Political Influence" (Advisor: Adam Bonica, Political Science; Second Reader: Margaret Levi, Political Science)
Abstract: As the importance of money to political success increases, along with the growing income gap, it is critical to consider new campaign finance reform efforts. I consider how a program that provides citizens with money to make political donations might be the key to drowning out the importance of big money. By allowing all citizens the opportunity for political influence through the act of donating, creating coalitions of financial support, and running for office, a Universal Basic Participation scheme might have a salubrious effect on American democracy. I specifically consider the benefits for historically underrepresented groups, and also explore Seattle's recent launch of a similar program.

1:20 Eric Zelikman, "Learning Is Its Own Reward: Exploring Worlds with Curiosity-driven Spiking Neural Networks" (Advisor: Nick Haber, Education; Second Reader: Tom Dean, Computer Science)
Abstract: While curiosity is an important part of human learning, it is rarely explicitly a part of machine learning: given some objective, machine learning algorithms optimize directly to get the best possible score. This has led to numerous impressive results but needs many examples and often fails in situations where performance feedback is limited. Approaches that attempt to incorporate curiosity generally still fall under this optimization paradigm. However, spiking neural networks (SNNs), which aim to better mimic biological neurons, present an alternative. Research has shown that local spike-timing-dependent learning rules allow for objective-free learning of complex patterns in data. However, limited work has been done on applying spiking neural networks in contexts where an agent must interact with a simulated world making a series of actions. We show that by rewarding a spiking neural network agent for the amount that it learns, our agent quickly learns to explore its environment in increasingly complex ways and seek increasingly novel situations. We demonstrate that, in terms of extrinsic rewards and various behavioral metrics, this approach often results in good performance much more quickly than other curiosity-driven methods.

1:30 Stephen Weyns, "Designing Embodied Dialogue Systems: A Way of Thinking about Character Conversation in Video Games" (Advisor: Ge Wang, Music: Second Reader: Chris Bennett, Computer Science) [video recorded]
Abstract: This thesis argues for a new way of thinking about designing engaging dialogue systems for video games. The proposed approach – called designing "embodied dialogue" systems – stands in contrast to and goes beyond designing traditional "descriptive" dialogue systems to create an intrinsically engaging moment-to-moment dialogue experience. On this basis, this work proposes three design principles for creating embodied dialogue systems, namely, such systems should be (1) expressive, (2) autonomy-inspiring, and (3) true to the underlying game genre. These principles are put into motion in a proof of concept game prototype called Embodied Dialogue; the ways in which the design approach shaped the game design are discussed in detail. This tangible instantiation of the proposed approach was qualitatively evaluated by play-testers as a way of assessing its initial effectiveness. Furthermore, the game itself serves as a critical exploration of embodied dialogue elements including strongly-timed audiovisual sonification of dialogue, audiovisual inner speech, and an expressive personality system. Overall, this thesis argues that designing embodied dialogue systems (and the various game elements that can be derived from using this approach) is a compelling way of thinking for creating more engaging video game dialogue systems.