BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Events - ECPv6.16.5.1//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:Events
X-ORIGINAL-URL:https://live-events-ucsc.pantheonsite.io
X-WR-CALDESC:Events for Events
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20250309T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20251102T090000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20260308T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20261101T090000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20270314T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20271107T090000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260709T133000
DTEND;TZID=America/Los_Angeles:20260709T153000
DTSTAMP:20260623T160412Z
CREATED:20260623T160248Z
LAST-MODIFIED:20260623T160412Z
UID:10014929-1783603800-1783611000@live-events-ucsc.pantheonsite.io
SUMMARY:Carrión\, H. (CSE) - Deep Learning Algorithms for Medical Image Representation Learning and Understanding
DESCRIPTION:AI-assisted clinical decisions in medicine\, and particularly in dermatology\, demand fine-grained understanding across diverse skin tones\, body sites\, and disease types\, yet expert-annotated datasets are scarce\, demographically imbalanced\, and almost devoid of rare presentations. This dissertation develops four deep learning systems for this low-label\, low-coverage regime. We introduce HealNet\, which learns wound healing stages from longitudinal photographs without any human labels\, reaching 90.6% downstream stage-classification accuracy on a small longitudinal cohort. The Fair\, Efficient\, and Diverse Diffusion (FEDD) model then leverages powerful diffusion-model embeddings to build a skin-tone-fair\, data-efficient classifier for skin lesions\, matching or exceeding state-of-the-art performance while using only 5-20% of available labels and contributing explicit skin-tone-stratified fairness evaluation of the work. Next\, Controllable Generation of Diverse Dermatological Imagery (cgDDI) re-tasks this diffusion model to controllably synthesize skin-tone-balanced dermatological imagery\, growing a small biopsy-confirmed dataset by over 400x and reaching state-of-the-art 90.9% accuracy and improved fairness in malignancy classification\, with a +13.9% cross-dataset gain on the Fitzpatrick17k benchmark. Finally\, we introduce D-Synth and DermDepth: a synthetic dermoscopic dataset with pixel-perfect 3D ground truth and a metric-scale foundation model that closes the loop into 3D dermatology\, correcting metric scale error from over 16x to under 1.1x on real dermoscopic data and enabling single-photograph measurement of lesion reconstruction: size\, area\, and volume without specialized hardware. All data\, code\, and models are released openly to support reproducibility and ongoing fairness research. \nEvent Host: Héctor Carrión\, Ph.D. Candidate\, Computer Science & Engineering \nAdvisor: Narges Norouzi \nZoom: https://ucsc.zoom.us/j/96678782408?pwd=71f0ObEnUMNgkZ9NYnpbFLMlg1Pdm0.1 \nPasscode: 0FMVtz
URL:https://live-events-ucsc.pantheonsite.io/event/carrion-h-cse-deep-learning-algorithms-for-medical-image-representation-learning-and-understanding-2/
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/png:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option-3.png
LOCATION:
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260710T110000
DTEND;TZID=America/Los_Angeles:20260710T123000
DTSTAMP:20260626T170310Z
CREATED:20260626T170310Z
LAST-MODIFIED:20260626T170310Z
UID:10014993-1783681200-1783686600@live-events-ucsc.pantheonsite.io
SUMMARY:Levine\, R. (CSE) - Validating GPU Memory Consistency and Safety at Scale
DESCRIPTION:Graphics Processing Units (GPUs) have become essential platforms for parallel computing\, supporting applications far beyond graphics. Central to GPU programming models is its memory consistency specification (MCS)\, which defines the semantics of concurrent shared-memory operations and interacts with other language features to determine security guarantees such as memory safety. Understanding whether implementations conform to an MCS\, and whether the MCS provides a sound abstraction of real hardware\, is essential for reasoning about GPU programs and validating implementations. \nThis thesis develops techniques and large-scale studies for validating GPU memory consistency and memory safety. First\, it introduces MC Mutants\, a mutation testing methodology that systematically evaluates GPU MCS test environments. Applied to WebGPU\, MC Mutants generates a suite of conformance tests and uncovers two implementation bugs. Next\, it presents GPUHarbor\, a browser- and Android-based framework for large-scale testing across commodity GPUs. GPUHarbor enables a study of 106 GPUs from seven vendors\, reveals two previously unknown memory consistency bugs\, and provides new insights into GPU behavior that inform subsequent architectural and security studies. Finally\, this thesis presents SafeRace\, a collection of security assessments and specification proposals for preserving WebGPU memory safety in the presence of data races. Evaluated across dozens of GPUs and 21 WebGPU compilation stacks\, SafeRace identifies vulnerabilities in multiple GPU implementations\, including one assigned a CVE\, and proposes a validated path toward stronger memory safety guarantees in WebGPU. \nEvent Host: Reese Levine\, Ph.D. Candidate\, Computer Science & Engineering \nAdvisor: Tyler Sorensen \nZoom: https://ucsc.zoom.us/j/94641390195?pwd=RWXp9aprCMqmaAo8nq7oKwqTt02zwN.1 \nPasscode: 628349
URL:https://live-events-ucsc.pantheonsite.io/event/levine-r-cse-validating-gpu-memory-consistency-and-safety-at-scale/
LOCATION:Engineering 2\, Engineering 2 1156 High Street\, Santa Cruz\, CA\, 95064
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/jpeg:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option-1.jpg
GEO:37.0009723;-122.0632371
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=Engineering 2 Engineering 2 1156 High Street Santa Cruz CA 95064;X-APPLE-RADIUS=500;X-TITLE=Engineering 2 1156 High Street:geo:-122.0632371,37.0009723
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260713T100000
DTEND;TZID=America/Los_Angeles:20260713T120000
DTSTAMP:20260707T160215Z
CREATED:20260707T160215Z
LAST-MODIFIED:20260707T160215Z
UID:10015010-1783936800-1783944000@live-events-ucsc.pantheonsite.io
SUMMARY:Scott\, J. (CSE) - Mechanistic Specialization Does Not Guarantee Performance: Evidence from Dual AttentionTransformers
DESCRIPTION:Dual Attention Transformers (DATs) extend decoder-only Transformers with a dedicated relational-attention stream\, making them a natural architecture for abstract identity rules such asABA and ABB. Surprisingly\, we find that comparably sized GPT-2 models outperform DATs on these tasks. We investigate this gap with two complementary mechanistic analyses. First\, causal mediation analysis shows that DATs exhibit stronger evidence of hypothesized symbolic mechanisms: symbol abstraction\, symbol induction\, and retrieval\, than GPT-2. Second\, a routing analysis shows why this specialization does not translate into better behavior: DATs make more wrong-copy errors\, can attend to the correct source token while still predicting the wrong token\, and show weak direct contribution from relational attention to the correct-versus-wrong outputmargin. Ablating positive-routing heads hurts performance\, while amplifying those headsimproves DAT more than matched controls. These results show that explicit relational attentioncan shape internal organization without guaranteeing task success. For identity-rule tasks\, performance depends not only on whether relational information is represented\, but whether it is routed to the final output position in a form that affects the next-token prediction. Because pretrained DAT and GPT-2 models differ in training data\, tokenizer\, and other implementation details\, these findings should be interpreted as evidence about the mechanisms used by existing models rather than as a definitive architectural comparison. Follow-up experiments will address these confounders through controlled training comparisons that match data\, scale\, and evaluation conditions across architectures. \nEvent Host: Jonathan Scott\, Ph.D. Student\, Computer Science & Engineering \nAdvisor: Leilani Gilpin \nZoom: https://ucsc.zoom.us/j/95404396322?pwd=0e0AegKSxhcFDDKrn08muHcqfHs6WW.1 \nPasscode: 985103
URL:https://live-events-ucsc.pantheonsite.io/event/scott-j-cse-mechanistic-specialization-does-not-guarantee-performance-evidence-from-dual-attentiontransformers/
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/jpeg:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option-1.jpg
LOCATION:
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260713T160000
DTEND;TZID=America/Los_Angeles:20260713T170000
DTSTAMP:20260708T155209Z
CREATED:20260708T155209Z
LAST-MODIFIED:20260708T155209Z
UID:10015011-1783958400-1783962000@live-events-ucsc.pantheonsite.io
SUMMARY:Kembay\, A. (ECE) - Sparse and Continual Foundations for Adaptive General Intelligence
DESCRIPTION:While the human brain learns continually\, mastering new tasks without forgetting\nthe old and adapting to unfamiliar ones from context alone\, modern neural networks\nstill lack both. To bridge the gap between biological adaptivity and modern AI\, we\nhave established foundational work on sparsity as a computational principle at three\nlevels of neural computation\, through salient feature masking that distills only the most\ninformative knowledge from a teacher\, quantized spiking neural networks whose sparse\nactivations mitigate catastrophic forgetting by updating weights only when new learn-\ning requires it\, and complex-pole value-path dynamics that give Transformer attention\na resonant\, positionally selective memory. Addressing the remaining bottleneck\, that\nthese sparse structures are fixed in advance rather than adapted to the task at hand\,\nwe propose a research roadmap centered on in-context meta-learning with sparse atten-\ntion priors\, enabling models to ‘learn to be sparse’ by inferring task-relevant structure\nfrom context alone\, without any weight update. Taken together\, this research seeks\nto unify brain-inspired sparsity with continual and in-context learning as a foundation\nfor adaptive general intelligence. \nEvent Host: Assel Kembay\, Ph.D. Student\, Electrical & Computer Engineering \nAdvisor: Jason Eshraghian \nZoom: https://ucsc.zoom.us/j/92202931005?pwd=peVIc4e03fUPwFqlGa6yWx6ZlL33lI.1 \nPasscode: 742766
URL:https://live-events-ucsc.pantheonsite.io/event/kembay-a-ece-sparse-and-continual-foundations-for-adaptive-general-intelligence/
LOCATION:Engineering 2\, Engineering 2 1156 High Street\, Santa Cruz\, CA\, 95064
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/png:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option-3.png
GEO:37.0009723;-122.0632371
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=Engineering 2 Engineering 2 1156 High Street Santa Cruz CA 95064;X-APPLE-RADIUS=500;X-TITLE=Engineering 2 1156 High Street:geo:-122.0632371,37.0009723
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260722T110000
DTEND;TZID=America/Los_Angeles:20260722T130000
DTSTAMP:20260708T160702Z
CREATED:20260708T160702Z
LAST-MODIFIED:20260708T160702Z
UID:10015012-1784718000-1784725200@live-events-ucsc.pantheonsite.io
SUMMARY:Holmes\, J. (CM) - Towards a Multi-dimensional Model of User Load
DESCRIPTION:Games user researchers (GURs) use various methods to understand when a game is overloading its players. In games research where data-driven multimodal approaches are necessary to drive insights\, the currently available tools to measure user load are coarse\, one-dimensional\, and often aggregated. The more dominant instruments\, such as the Cognitive Load Scale (CLS) and the NASA-TLX\, rely on player reflections of mental effort\, primarily at the end of the playtest session\, to distinguish different cognitive load types. This makes it difficult to: (1) understand where specifically players are struggling and experiencing high load\, especially at the non-reflective subconscious level\, (2) identify where that load is primarily coming from (e.g.\, perceptual clutter or difficulty/skill imbalance)\, and (3) examine user overload at scale\, a crucial component of designing a game with large player bases. Telemetry is the behavioral record of what players are doing from moment to moment\, in varying degrees of granularity. Telemetry has served as a powerful tool to understand player behaviors at scale\, yet is rarely used to measure user load\, especially through a validated multidimensional framework. This dissertation proposes that behavioral signatures of specific load constructs are observable in game telemetry\, and a model built and validated on such telemetry can measure the distinct components of each load at the moment-to-moment granularity of individual play\, as opposed to aggregated magnitude. This dissertation consists of three parts: (1) Validation through construct manipulation and reference measurement. Specifically\, manipulating theoretically grounded load constructs and confirming that the proposed telemetry features respond as predicted\, relative to the established measurements collected alongside them (e.g.\, NASA-TLX\, pupillometry\, secondary-task). (2) Individual-level validation through rigorous longitudinal examination of the same players repeatedly across many sessions such that load constructs can be tracked at the within-person granularity. This is necessary to establish that the measure works for an individual player and not just for population averages. (3) Test the user load model by applying it to naturalistic game telemetry. Additionally\, this phase will entail the development of an insight-oriented measurement tool for GURs based on our validated user load model. The overarching contribution is a behavioral\, telemetry-based method for measuring multidimensional user load in games\, validated to measure load within each person (individual-level). This gives GURs a scalable tool and replicable process for detecting user load in commercial game telemetry. \nEvent Host: Jonattan Holmes\, Ph.D. Student\, Computational Media \nAdvisor: Magy Seif El-Nasr \nZoom: https://ucsc.zoom.us/j/98245962806?pwd=HnkwPMFSamQJFrE5aihbZbKDBbt4s9.1 \nPasscode: 347521
URL:https://live-events-ucsc.pantheonsite.io/event/holmes-j-cm-towards-a-multi-dimensional-model-of-user-load/
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/jpeg:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option2.jpg
LOCATION:
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20260723T120000
DTEND;TZID=America/Los_Angeles:20260723T140000
DTSTAMP:20260708T162027Z
CREATED:20260708T162027Z
LAST-MODIFIED:20260708T162027Z
UID:10015013-1784808000-1784815200@live-events-ucsc.pantheonsite.io
SUMMARY:Li\, J. (CM) - Detecting Failure to Adapt: Reading Self-Regulated Learning Breakdowns from Game Telemetry through Plan Recognition
DESCRIPTION:Three learners who fail the same level of an educational game the same number of times can be failing in three different ways\, and the difference determines what each should do next. Yet the measures a game’s logs are usually reduced to (completion time\, error counts\, mastery estimates) render the three identical. This proposal takes one breakdown as its object: failure-to-adapt\, the case where the game has repeatedly surfaced evidence that a learner’s current approach is failing and the learner’s approach shows no responsive change. The construct is grounded in Winne and Hadwin’s monitor-and-control model of self-regulated learning and defined at the level of the learner’s plan. To detect it\, a plan-recognition engine maintains a continuously updated probability estimate of which strategy the learner is executing across the whole trace; an episode is flagged when that estimate shows no evidence-responsive revision. Because behavior alone cannot settle what broke down\, flagged episodes are validated against learners’ own verbal reports\, coded blind\, and decomposed into monitoring failure\, control failure\, or control the trace cannot show. Three studies carry the work: detection and diagnosis on real telemetry from an educational game\, including a comparison against the analytics the field already runs; a formative study of what a facilitator (an instructor or teaching assistant running a class play session) must see to judge correctly which learners need attention; and a documented authoring case carrying the detection to a second game. The contribution is knowledge for game-based-learning researchers: a theory-grounded construct\, a validated way to detect it from play\, and the authoring knowledge to embed that detection in new games. \nEvent Host: Jiahong Li\, Ph.D. Student\, Computational Media \nAdvisor: Magy Seif El-Nasr \nZoom: https://ucsc.zoom.us/j/93238603235?pwd=zENRsu82HRj4JYKcMEn9MZibU8kC7F.1 \nPasscode: 835328
URL:https://live-events-ucsc.pantheonsite.io/event/li-j-cm-detecting-failure-to-adapt-reading-self-regulated-learning-breakdowns-from-game-telemetry-through-plan-recognition/
CATEGORIES:Ph.D. Presentations
ATTACH;FMTTYPE=image/jpeg:https://live-events-ucsc.pantheonsite.io/wp-content/uploads/2026/04/ph.d.-presentation-graphic-option-1.jpg
LOCATION:
END:VEVENT
END:VCALENDAR