Aaron Sun

I'm a PhD Student at UMass Amherst studying Computer Vision under Subhransu Maji, currently focusing on improving identity-preserving generation for scientific accuracy on fine-grained domains.

Email  /  CV  /  Scholar  /  Github  /  LinkedIn

Research

We propose WildProp, a pipeline which uses unconstrained photographs to measure population-level body proportions at global scale.

We introduce the NABirds Look-Alikes (NABLA) dataset, a dataset for identity-preserving generation for birds, and show how training on proxy identity pairs improves performance on seen and unseen species.

We introduce the RealBirdID benchmark, which focuses on quantifying and evaluating uncertainty for encoder and VLM models on bird species.

We present a detailed analysis comparing the efficiacy of different molecule representations and training strategies on zeolites.

We introduce the L48 Spectrogram dataset, a fine-grained, real-world benchmark for SPML learning, and propose a new regularization method which improves existing techniques.

We present a detailed analysis comparing the efficiacy of different molecule representations and training strategies on zeolites.

We present a method of task representation using box embeddings and demonstrate how they are able to capture asymmetric relationships between tasks and generalize to unseen tasks.

We present a representation learning framework using CNNs and 3D volumetric representations for predicting adsorption in zeolites.

We present a set of metrics that utilize vision priors to effectively assess the performance of saliency methods on image classification tasks.