Publications

Naming systems trade off informativeness against complexity. Focusing on kinship naming, we revisit two simplifying assumptions of prior work, universal communicative need and optimal listeners, using data from four languages. Adopting a referential game setup from emergent communication, we show that trade-off optimality is not only theoretically achievable but also emerges empirically in learned communication systems.

Most hyperbolic vision architectures are only extrinsically hyperbolic, falling back to Euclidean computation before re-projecting onto the manifold, a shortcut shown to undercut the benefits of hyperbolic geometry. We introduce LoViT, the first Intrinsic Lorentz Vision Transformer, carrying intrinsic treatment throughout, including a residual connection defined as geodesic interpolation. Trained from scratch, LoViT surpasses a Euclidean baseline and extrinsic hyperbolic transformers on CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Published in TMLR (2025). Selected for presentation at the 2025 Machine Learning Reproducibility Challenge held at Princeton University. This study evaluates the claims and results of Are Your Models Still Fair? Fairness Attacks on Graph Neural Networks via Node Injections. Assessed implementation, experimental setup, and generalizability to provide insights into the robustness of the original findings.