UC Santa Barbara

Ambitious presumptivism says that all our testimony based beliefs are on-balance immediately and defeasibly warranted. The rational deception objection says that ambitious presumptivism is not true because it is sometimes rational for a speaker to assert lies rather than truths. One logically possible reply is to argue that it is never rational for a speaker to assert lies rather than truths. In this essay, I develop such a non-conciliatory response to the rational deception objection.

In chapter 1, I explain ambitious presumptivism and the rational deception objection. I identify Kant's prohibition against lying as a historical predecessor to the non-conciliatory response to the rational deception objection. I then identify Burge as the heir apparent to a neo-Kantian non-conciliatory response to the rational deception objection. In chapter 2, I explain my interpretation of how Burge is heir apparent to a neo-Kantian non-conciliatory response. I call Burge's response the ``functional unity'' argument.

In chapter 3, I defend my attribution to Burge of the functional unity argument. In chapters 4 and 5, I defend the functional unity argument itself from the most influential objections raised against it. in chapter 6, I defend the functional unity argument from the classic murderer-at-the-door objection that dogged Kant's prohibition against lying.

This mixed methods study explored how individuals learn about emotional and sexual intimacy, how individuals communicate about intimacy in their romantic relationships, and what socialization factors facilitate or impede sexual communication in relationships. Three theoretical frameworks – attachment theory, family communication patterns theory, and adverse childhood experiences – were used to quantitatively test how socialization factors affect individuals’ communication and satisfaction in current relationships. Romantic couples (N = 135 dyads) discussed emotional and sexual intimacy in their relationships during a recorded Zoom conversation and completed online surveys before and after the Zoom conversation. A subsample of original participants (n = 31) completed follow-up interviews that went more in depth about their socialization of intimacy and their perceptions of their relationship following the Zoom conversation. Multiple methods were used to analyze the data. Actor-partner interdependence structural equation modeling tested the hypothesized quantitative models. Attachment avoidance and anxiety had the strongest effect on participants’ current relational communication. However, both partners’ communicative responsiveness, fear of emotional intimacy, and general sexual communication affected each other’s relationship and sexual satisfaction, thus demonstrating the importance of dyadic analyses for interpersonal relationships. A phronetic iterative analysis (Tracy, 2020) was used to analyze the conversations and follow-up interviews for overarching themes. The qualitative analyses of both the conversation and interview data produced four themes: (a) Socialization of Intimacy, (b) Learning About Intimacy is a Continuous Process, (c) Intimacy Displays are Either Modeled or Compensated, and (d) Emotional and Sexual Intimacy Build Over Time. A fifth theme gleaned from the interview data, Conversation Served as an Intimacy Intervention, illuminated that the Zoom conversation functioned as a means for many couples to have subsequent intimacy conversations after participating in the study. The findings from this study have pragmatic implications for all couples, especially those who struggle with discussing intimacy. The findings from this study are also useful for clinicians who work with individuals and/or couples to better understand their communication and relationships by examining how their past experiences affect their current communication patterns.

Numerous large vision models for natural images, such as SAM, Florence-2, and GPT-4, have achieved state-of-the-art (SOTA) performance, largely due to vast amounts of image and text data available online. Smaller models like EfficientSAM and CLIP have also shown the potential of achieving significant results with comparatively less data. However, real-world scientific problems, particularly in remote sensing, present unique challenges due to the complexity of the data and scarcity of annotations. These problems often require data from multiple sources, such as hyperspectral sensors on airplanes and multispectral sensors on satellites, which are expensive and time-consuming to acquire.This dissertation addresses the key question: how can large vision models be built and trained effectively under data constraints? The proposed solution involves integrating domain-specific knowledge into large vision models, specifically vision transformers, to optimize their performance and training efficiency. By incorporating core signal processing techniques, domain-specific knowledge is encoded as prior information, guiding the feature extraction process and refining randomly initialized queries via a query refiner module. This approach accelerates convergence with limited training data. Three key applications are explored: (1) methane detection in remote sensing from aerial imagery, (2) animal detection and classification in large grasslands for ecological studies, and (3) estimation of physiological signals such as ECG and ISTI for stress assessment in biomedical contexts. This research establishes an optimal methodology for embedding domain-specific knowledge into deep learning models, thereby enhancing performance in data-limited environments. It provides valuable insights for improving the applicability of vision transformer-based models across various domains, contributing to advancements in computer vision research and its practical real-world applications.

The rapid advancement of deep learning has significantly impacted the medical domain, benefitting various applications including clinical decision-making, personalized treatment, and medical education. Deep learning applications in the medical domain can be categorized based on the data types used: 1) Numerical measurements modeling: building models on numerical clinical measurements, including static and time-series data; 2) Natural Language Processing (NLP): training models on medical textual data such as doctor-patient conversations and clinical notes; and 3) Multimodal learning: leveraging data from multiple modalities to enhance the model's medical capacity and performance. This thesis presents works in these three categories, aiming to advance AI systems that can assist clinicians in enhancing healthcare outcomes and efficiency.

In numerical measurements modeling, despite the effectiveness of deep learning models in decision support, many studies rely on extensive public datasets, overlooking the data scarcity in small hospital settings. We address this by utilizing domain adaptation techniques to improve modality prediction in ICU patients with limited data.

Concerning NLP, while Large Language Models (LLMs) like ChatGPT and GPT-4 have shown promising results, privacy concerns restrict their direct use in healthcare. We propose integrating medical knowledge from LLMs into local models for decision support to alleviate these privacy concerns. Furthermore, instruction tuning has become crucial in aligning LLMs with human intents and has shown potential in medical applications. However, existing medical LLMs ignore the diversity of tuning data, limiting their ability to follow medical instructions and generalize. This thesis presents a novel approach to generating a diverse, machine-generated medical instruction-following dataset and demonstrates that the model tuned on this dataset achieves superior performance in both medical and general domains.

For multimodal learning, although improvements have been seen in medical predictions using multimodal data, challenges in modeling irregularities within each modality and integrating irregular time information into the multimodal representation persist. We introduce strategies for addressing these challenges in multimodal electronic health records to enhance predictions for ICU patients.

Finally, we summarize the key findings and discuss future research directions to push the boundaries of deep learning in medical applications.

Recent works have demonstrated that New Zealanders who are frequently exposed to Māori in everyday life, but do not speak it, have an extensive memory store of Māori forms, called a proto-lexicon (Oh et al., 2020). This proto-lexicon is composed of morphs - words and word pieces that recur with statistical regularity in language usage that are learned through statistical learning (Ngon, et al., 2013). The proto-lexicon endows Non-Māori-Speaking New Zealanders (NMS) with rich implicit knowledge of Māori, which permits them to morphologically segment Māori words at above-chance levels (Panther et al., 2023a). Prior works (Saffran et al., 1996; Saffran 2003; Frank et al., 2013) have shown how statistical learning helps in implicit learning, but only in artificial languages. Oh et al. (2020) is one of the first studies to have shown this in real world exposure. In this work we use Morfessor (Smit et al., 2014), an unsupervised Bayesian segmentation model that identifies statistically recurrent morphs across words under the assumption of morphological concatenativity, to build on these recent studies to investigate the extent of statistical learning used by NMS. We use Morfessor as our control statistical learner to perform two analyses. In our first analysis, we compare NMS and Morfessor to an expert Māori Speaker’s (MS) ability to segment words into morphs. Comparing NMS and Morfessor’s segmentation performances, we show the differences and similarities in the segmentation and learning process, and how it is affected by the statistical properties of the language. Further, using an error analysis on the segmentations, we gain insights into their underlying assumptions used in their segmentation process. The results of analysis 1 suggest that NMS may be sensitive to more than Morfessor, e.g. templates. As a follow up to these results, in our second analysis, we dive deep into the results of the concatenative category of words whose structure closely resembles Morfessor’s assumption. By generating pseudo-Māori words for this category and testing Morfessor’s performance on them, we provide insights into how the statistical learning of real Māori morphs depends on explicit cues which it does not have access to – which the NMS seem to have some access to, where in they use the statistical regularities by taking a templatic approach in order to segment the words into morphs. The most recent updated version of this work for publication can be found here : http://arxiv.org/abs/2403.14444.

The field of machine learning has seen great strides in the past two decades, leading to groundbreaking advancements in various domains. Yet many challenges remain in AI research. Scalability remains a persistent hurdle, with current models struggling to efficiently manage increasing amounts of data and complexity. The challenge of scalability is further highlighted by the power of the human brain, which, with its remarkable efficiency, underscores the limitations of current machine learning models. Finally, both industry and academia have faced considerable difficulties in advancing robotics, particularly in enabling neural networks to interact effectively with the physical world.

In this work, we explore how the lens of dynamical systems can address all three of these challenges. We tackle memory constraints by framing a quantization algorithm as a fixed-point equation, enabling efficient differentiation. We investigate an energy-based, biological model of human memory and reinterpret the widely used self-attention mechanism through this model. Lastly, we leverage contraction theory to train a neural network that can follow a trajectory with stability and robustness. In the first two problems, we utilize dynamical systems to differentiate neural networks, while in the latter, we employ neural networks to learn a dynamical system. Through our understanding of dynamical systems, we build on theoretical advancements and practical applications in AI, offering new insights into both memory optimization and robust robotic manipulation.

Deep neural networks have demonstrated significant achievements across various fields, yet their memory and time complexities present obstacles for implementing them on resource-constrained devices. Compressing deep neural networks using tensor decomposition can decrease both memory usage and computational costs. The performance of a low-rank tensorized network depends on the choices of hyperparameters including the tensor rank and geometry. Previous studies have concentrated on identifying optimal tensor ranks. This thesis studies the effect of tensor geometry used for folding data for low-rank tensor compression. It is demonstrated that tensor geometry significantly affects compression efficiency of the tensorized data and model parameters. Consequently, a novel mathematical formulation is developed to optimize tensor geometry. The tensor geometry optimization model is adopted for efficient deployment of low-rank neural networks. The presented tensor geometry optimization model is combinatorial and thus challenging to solve. Therefore, surrogate and relaxed versions of the model are developed and various methods including integer linear programming, graph optimization, and random search algorithms are applied to solve the presented optimization model. The proposed tensor geometry optimization achieved a notable reduction in both the memory and time complexities of neural networks while maintaining accuracy. The developed methods can be applied for hardware-software co-design of artificial intelligence (AI) accelerators particularly on resource-constrained devices.

A more complete understanding of the biological visual system can inspire the design of computer vision algorithms, and building accurate models constitutes an important step to such an understanding. We utilize computational and deep learning approaches to close the gaps in the literature on modeling the retina and the primary visual cortex (V1), the two important components of the early visual processing pathway.Firstly, to address the lack of a comprehensive computational model of retinal degeneration, we present a biophysically detailed model of the cone pathway in the retina that simulates responses to light and electrical stimulation. Anatomical and neurophysiological changes due to retinal degenerative diseases were systematically introduced. The model was not only able to reproduce common findings about retinal ganglion cell (RGC) activity in the degenerated retina, but also offered testable predictions about the underlying neuroanatomical mechanisms. These insights may further our understanding of retinal processing and inform the design of retinal prostheses. Secondly, to argue for more emphasis on freely moving experimental design, we propose an analysis of the retinal input during free exploration in mice. Mice were able to employ compensatory and gaze-shifting eye-head movements to sample the visual environment during natural locomotion. We found that eye movements preferred features such as edges and textures. A deep learning predictive model of gaze shifts indicated that the upper peripheral visual field contributed most to the prediction, consistent with animal behavior such as predator detection. These results may provide implications for visual processing beyond head-fixed preparations. Lastly, to bridge the gap in predictive modeling tailored for neural data gathered from freely moving experimental paradigms, we introduce a multimodal recurrent neural network that integrates gaze-contingent visual input with behavioral and temporal dynamics to explain V1 activity in freely moving mice. The model achieves state-of-the-art predictions of V1 activity during free exploration. Analyzing our model using maximally activating stimuli and saliency maps, we reveal new insights into cortical function, including the prevalence of mixed selectivity for behavioral variables in mouse V1. Our model offers a comprehensive deep-learning framework for exploring the computational principles underlying V1 neurons in freely-moving animals engaged in natural behavior.

Efforts to improve gas turbine fuel efficiency face a significant obstacle in the degradation of thermal and environmental barrier coatings (T/EBC) made of refractory oxides. Although these coatings effectively protect against heat and water vapor, they inadequately address the challenges posed by molten silicates made of engine-ingested ash, sands, and dusts. Additionally, certain candidate coating materials are susceptible to microcracking driven by anisotropy, compromising their ability to prevent ingress of gas species that hasten component failure. This dissertation features new computational and experimental approaches to investigate and design T/EBC materials, offering fundamental insights into molten silicate attack and grain boundary microcracking.

The first aspect of this dissertation addresses the need to develop a protocol for selecting compositions of silicate deposits for the purpose of assessing performance of candidate coating materials. The crux of the study is the statistical analyses (i.e. principal component analysis and \textit{k}-means clustering) of a curated database to identify exemplary compositions. Potential chemical interactions between the exemplary silicates and several coating materials were explored using thermodynamic calculations. The computational framework is expected to inform future work on investigating potential attributes and deficiencies of new coating materials.

The second aspect of this dissertation explores the potential of HfO$_2$-based coating materials as barriers to molten silicates. (These materials are attractive candidates because of their high-temperature phase stability and resistance to water-vapor mediated volatilization.) Since pure HfO$_2$ is susceptible to extensive grain boundary microcracking, two alternatives were explored: one based on crack-free hafnia/hafnon composites, and one based on hafnate compounds, building on the understanding of corresponding zirconate compounds.

High-temperature exposures of the hafnia/hafnon composites to two exemplary silicates revealed that in all cases interactions include extensive grain boundary penetration, presumably driven by the lack of chemical equilibrium between the composite constituents and the silicate melts. The findings prompted a comparative study of the interactions of Gd-hafnate and Gd-zirconate with the exemplary silicate melts. As with the hafnia/hafnon composites, the melts readily penetrate the Gd-hafnate, but not the Gd-zirconate, owing to rapid co-formation of protective barrier layers of apatite and fluorite above the zirconate. Further insights on the underlying cause of these differences were gleaned from complementary studies in which short exposures (1-3 min) using Gd-lean versions of the compounds were used to quantify the kinetics of the interactions. The results showed that Hf$^{4+}$ diffuses in the melt more slowly than Zr$^{4+}$. Further, the findings implied that the dissolution processes are diffusion-controlled and that the slower kinetics of the processes as well as the subsequent crystallization necessary for protection of the hafnia-based systems may be the root cause of their poor resistance to silicate attack. More broadly, the work demonstrates how fundamental studies of the kinetics of dissolution, diffusion, and crystallization can inform coating material selection and assessment.

The third aspect of this dissertation responds to a longstanding need for high throughput assessments of the driving forces for grain boundary microcracking in brittle materials of high anisotropy. The framework is based on a finite element approach to computing the energy release rate (ERR) for intergranular cracking. Simulations of bicrystals and polycrystals comprising periodic arrays of hexagonal grains were performed for 35 materials covering all 7 crystal systems. The assessments reveal that, while crystal system is not a determinant of likelihood for cracking, materials with large thermal and elastic anisotropy are more sensitive to grain orientations. Moreover, ERR distributions for bicrystals and polycrystals of the same material are in reasonable agreement with one another, implying that the details of neighboring grains in the polycrystals are relatively unimportant with regard to the average crack driving forces. Therefore, future studies might avoid the computational cost of simulating large polycrystals by opting for bicrystals alone, with the recognition that the details of the tails of the ERR distributions may not be accurately depicted. From a broader perspective, the high throughput nature of the approach should find considerable utility in not only the design of monolithic and multi-phase coatings, but also patterned and textured materials that avoid microcracking.

This dissertation consists of three essays in applied microeconomics. While the topics vary, the three papers are united in their use of causal inference techniques and their relevance to policy: each paper either evaluates effects of an existing policy or examines whether new policies are needed for consumer protection.

The first essay examines the effects of access to Buy Now, Pay Later (BNPL) on financial well-being. Many American consumers have limited access to credit, raising the question of whether an increase in credit access would make them better off. Fully rational individuals would use an increase in credit access to smooth consumption, yet real consumers may make financial mistakes by accumulating debts they cannot repay. I study the effects of making BNPL accessible to American consumers, including those who otherwise have limited access to credit. This paper provides the first causal evidence of how access to BNPL affects severe measures of financial distress and credit scores. Using credit bureau data and a two-way fixed effects identification strategy that exploits geographic and temporal variation in availability of BNPL at a large retailer, I find that access to BNPL reduces financial distress arising from late or missed debt payments. The total amount past due decreases by 2.4% and the number of current delinquencies decreases by 0.2%. Heterogeneity analysis reveals that these effects are strongest among consumers with “fair” credit scores, the second-lowest credit score category. I also find that BNPL access increases credit scores by an average of 1.6 points and increases use of non-BNPL credit. These results suggest that access to BNPL reduces financial distress rather than causing consumers to accumulate unsustainable debts.

The second essay studies how public financing for political campaigns affects political participation and campaign contributions. Seattle’s Democracy Vouchers program provides a unique form of public financing for political campaigns in which voters decide how to allocate public funding across candidates. This paper is the first to study the effects of public financing for political campaigns on political participation. I estimate that the Democracy Vouchers program increases voter turnout by 4.9 percentage points, suggesting that public financing programs can increase political participation. I also find that campaigns become more reliant on small contributions. For city council candidates, dollars from small contributions under $100 increase by 156% while dollars from large contributions over $250 decrease by 93%.

The third essay examines how legalizing marijuana affects fertility. State-level marijuana legalization has unintended consequences, including its effect on fertility. Marijuana use is associated with behaviors that increase fertility as well as physical changes that lower fertility. In this paper, I provide the first causal evidence of the effects of recreational marijuana legalization on birth rates using a difference-in-differences design that exploits variation in marijuana legalization across states and over time. The main result is that legalizing recreational marijuana decreases a state’s birth rate by an average of 2.78%. Heterogeneity analysis shows that the largest decrease in the birth rate occurs among women close to the end of their child-bearing years. I find suggestive evidence of increases in days of marijuana use per month and in the probability of being sexually active. Together, these findings show that the physical effects of marijuana use have the dominant effect on fertility. Finally, I examine the effects of medical marijuana legalization on fertility and find a smaller, statistically insignificant decrease in the birth rate, which is consistent with the smaller increase in marijuana use that results from medical legalization.