BEGIN:VCALENDAR
VERSION:2.0
METHOD:PUBLISH
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:STANDARD
TZNAME:GMT
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
BEGIN:DAYLIGHT
TZNAME:BST
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:6a0b7500151e1
DTSTAMP:20260518T202224Z
DTSTART;TZID=Europe/London:20260401T160000
DTEND;TZID=Europe/London:20260401T170000
TZID:Europe/London
SUMMARY:Computer Science PGR Seminar
DESCRIPTION:Weiye Li will present Deep Generative Model in the Regime of Incomplete Data    Abstract: Deep generative models offer a principled probabilistic framework for estimating true data distributions, P(X). Specifically, Variational Autoencoders (VAEs) map complex data into a structured latent space, z, which serves as a highly useful representation for downstream tasks. However, the quality of this learned representation is heavily dependent on the completeness of the input data. In real-world machine learning settings, features are frequently corrupted or missing, which presents a fundamental challenge to training deep generative models.    This research investigates the robust learning of deep generative models under missingness. We first demonstrate that missing features induce greater uncertainty in the latent space. This is illustrated qualitatively by multi-modal posteriors, and quantitatively by increased encoder entropy compared to complete inputs. Because standard VAEs enforce a unimodal Gaussian assumption on the latent space, they struggle under this ambiguity, often collapsing onto the wrong mode or averaging across modes to produce blurry outputs. Rather than artificially enriching the VAE posterior with mixtures or normalizing flows, we propose CMissVAE, a framework that explicitly conditions on auxiliary labels to resolve multi-modal ambiguity, rendering the inference problem manageable within a standard unimodal setup.    Xinya Gong will present Sensor-to-Text Alignment for Human Motor Behaviour: Towards Interpretable Parkinson's Disease Assessment    Abstract: This work explores whether human movement data can be translated into natural language.    Rather than predicting disease labels, it aims to bridge sensor signals and interpretable text. While recent multimodal approaches have aligned signals with language, existing methods rely on predefined concepts or coarse statistical descriptions, leaving fine-grained motor dynamics insufficiently grounded in meaningful semantics.    Using handwriting data in Parkinson's disease as a case study, this work investigates how writing dynamics---such as speed, pauses, and oscillations---can be mapped into concept-grounded textual descriptions. We adopt a signal-concept-text formulation, where weakly supervised motor concepts derived from kinematic features are encoded as text prototypes and aligned with signal representations via a lightweight alignment module. https://events.st-andrews.ac.uk/events/computer-science-pgr-seminar-8/
LOCATION:Jack Cole Building, North Haugh, KY16 9SX, St Andrews, Fife, Scotland
URL:https://events.st-andrews.ac.uk/events/computer-science-pgr-seminar-8/
End:VEVENT
End:VCALENDAR