Simon Fong

Achieve more factual radiology reports by unifying neuro-symbolic reasoning with active uncertainty minimization, outperforming existing encoder-decoder and retrieval-augmented methods.

SubQuad tackles the computational bottleneck of adaptive immune repertoire analysis with a near-quadratic-free approach that also corrects for dataset imbalances, enabling more equitable and scalable biomarker discovery.

By adversarially synthesizing graph structures and self-correcting node labels, AdvSynGNN achieves state-of-the-art robustness against structural noise and heterophily in graph neural networks.

LiveGraph tackles the cold-start problem in exercise recommendation by actively constructing a graph that connects active and inactive students, leading to improved accuracy and diversity.

Make your multimodal models immune to missing modalities with a training framework that selectively collapses modality information, boosting robustness without sacrificing performance.

Achieve practical privacy-utility trade-offs in multimodal sentiment analysis with Missing-by-Design (MBD), a framework that surgically removes modality-specific information without full retraining.

Achieve robust multimodal sentiment analysis, even with noisy or incomplete data, by explicitly modeling hierarchical relationships in hyperbolic space.

By actively exploring knowledge graphs with a differentiable neural-symbolic approach, NeuroSymActive achieves strong KGQA accuracy while drastically reducing the computational cost of graph lookups.

Achieve a better trade-off between search precision and carbon footprint with GaiaFlow, a framework that uses semantic-guided diffusion tuning to make neural search systems more sustainable.

Generate expansive, high-fidelity 3D environments with Gaussian Splatting using significantly less data by intelligently sampling viewpoints and hallucinating details with diffusion models.