Matthew Gwilliam

I am a fourth year Ph.D student in the department of Computer Science at the University of Maryland (UMD), advised by Professor Abhinav Shrivastava. I am studying computer vision.

I completed my B.S. in Computer Science at Brigham Young University in 2019. During undergrad I worked part-time at Qualtrics, and after graduation I worked there full-time before I started my Ph.D in 2020.

While at BYU, I was fortunate to work with Ryan Farrell, who helped me grow as a researcher and a person and ultimately decide to pursue my graduate degree.

Email  /  CV  /  Google Scholar

I am interested in computer vision models that learn without labels. More specifically, I am interested in methods that can learn universal image representations in an unsupervised manner. Currently, that work focuses on models based on diffusion and implicit neural representation (but mostly on INR). I am working with the sorts of tasks that these models are useful for: video retrieval, compression, generation; image classification, clustering, etc.

XINC dissects Implicit Neural Representation (INR) models to understand how neurons represent images and videos and to reveal the inner workings of INRs.

Explore diffusion models as unified unsupervised image representation learning models for many recognition tasks. Propose DifFormer and DifFeed, novel mechanisms for fusing diffusion features for image classification.

Propose an alternative to paragraphs for long video retrieval, specifically, the 10k Words problem: every video should be able to be matched with any valid description, so we generate many descriptions for every video (for 3 long video datasets), evaluate accordingly, and introduce novel finetuning for better performance.

Show the potential of diffusion models as unified unsupervised image representation learners.

Combine the strengths of implicit (NeRV) and explicit (autoencoder) representation to create a hybrid neural representation for video with good properties for representation, compression, and editing.

Make implicit video representation networks generalize to unseen data by swapping time embedding for content-aware embedding that is computed as a unique summary of each frame.

Examine, compare, and contrast popular unsupervised image representation learning methods, showing that there are significant differences based on specific algorithm used, and "supervised vs. unsupervised" comparisons which neglect these differences tend to over-generalize.

Reveal the role of data in racial bias for facial recognition systems, and the flaws underlying the assumption that balanced data results in fair performance.

Uncover the large, often-ignored amount of variance in FGVC systems across training runs, on the dataset level, but more particularly in terms of the classification performance for individual classes.

Propose smart practices to optimize image curation, such that classification accuracy is maximized for a given constrained dataset size.

Template credits