<mods:mods xmlns:mods="http://www.loc.gov/mods/v3" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.loc.gov/mods/v3 http://www.loc.gov/standards/mods/v3/mods-3-7.xsd"><mods:titleInfo><mods:title>Advancing Attention Mechanisms in Multimodal Deep Learning Models</mods:title></mods:titleInfo><mods:typeOfResource authority="primo">dissertations</mods:typeOfResource><mods:name type="personal"><mods:namePart>Golovanevsky, Michal</mods:namePart><mods:role><mods:roleTerm type="text">creator</mods:roleTerm></mods:role></mods:name><mods:name type="personal"><mods:namePart>Singh, Ritambhara</mods:namePart><mods:role><mods:roleTerm type="text">Advisor</mods:roleTerm></mods:role></mods:name><mods:name type="personal"><mods:namePart>Eickhoff, Carsten</mods:namePart><mods:role><mods:roleTerm type="text">Reader</mods:roleTerm></mods:role></mods:name><mods:name type="personal"><mods:namePart>Sun, Chen</mods:namePart><mods:role><mods:roleTerm type="text">Reader</mods:roleTerm></mods:role></mods:name><mods:name type="corporate"><mods:namePart>Brown University. Department of Computer Science</mods:namePart><mods:role><mods:roleTerm type="text">sponsor</mods:roleTerm></mods:role></mods:name><mods:originInfo><mods:copyrightDate>2026</mods:copyrightDate></mods:originInfo><mods:physicalDescription><mods:extent>xii, 98 p.</mods:extent><mods:digitalOrigin>born digital</mods:digitalOrigin></mods:physicalDescription><mods:note type="thesis">Thesis (Ph. D.)--Brown University, 2026</mods:note><mods:genre authority="aat">theses</mods:genre><mods:abstract>Multimodal models are central to modern Artificial Intelligence systems, with growing relevance in domains&#13;
such as healthcare. Their success is often credited to attention mechanisms, which enable rich cross-modal&#13;
interactions. However, attention scales poorly when more than two modalities are integrated, and its internal&#13;
functions remain poorly understood. This thesis addresses the scalability and interpretability of attention&#13;
in multimodal models by introducing novel integration mechanisms, interpretability pipelines, and targeted&#13;
interventions. In the first chapter, I introduce One-Versus-Others attention, a scalable alternative to self- and&#13;
cross-attention that reduces computational cost while preserving accuracy in high-modality settings such as&#13;
clinical data integration. In the second chapter, I introduce NOTICE, a vision–language causal mediation&#13;
framework that enables the discovery of human-interpretable functions in attention mechanisms, revealing&#13;
distinct roles for cross- and self-attention. In the third chapter, I analyze the internal competition between&#13;
memorized priors and new perceptual input in multimodal models, demonstrating that steering vectors can&#13;
reallocate attention to favor either prior knowledge or new visual evidence. In the final chapter, I examined&#13;
how specific groups of attention heads induce prompt-copying behavior in vision–language models, leading&#13;
to systematic prompt-induced hallucinations on numerical and semantic tasks. Through targeted ablation, I&#13;
showed that these hallucinations can be reduced without retraining, and that the mechanisms by which copying&#13;
is suppressed differ across models while producing a shared shift toward visual grounding. The findings&#13;
presented in this thesis deepen our understanding of attention in multimodal models, offering concrete tools&#13;
for scalability and interpretability. More broadly, they establish foundations for building multimodal systems&#13;
that are trustworthy and reliable in real-world applications.</mods:abstract><mods:subject authority="fast" authorityURI="http://id.worldcat.org/fast" valueURI="http://id.worldcat.org/fast/01034365"><mods:topic>Natural language processing (Computer science)</mods:topic></mods:subject><mods:subject authority="fast" authorityURI="http://id.worldcat.org/fast" valueURI="http://id.worldcat.org/fast/00817247"><mods:topic>Artificial intelligence</mods:topic></mods:subject><mods:subject authority="fast" authorityURI="http://id.worldcat.org/fast" valueURI="http://id.worldcat.org/fast/02032663"><mods:topic>Deep learning (Machine learning)</mods:topic></mods:subject><mods:language><mods:languageTerm authority="iso639-2b">English</mods:languageTerm></mods:language><mods:recordInfo><mods:recordContentSource authority="marcorg">RPB</mods:recordContentSource><mods:recordCreationDate encoding="iso8601">20260427</mods:recordCreationDate></mods:recordInfo></mods:mods>