Angus Galloway
PhD Student in Systems Engineering
University of Guelph
Vector Institute
Ontario, Canada
My research aims to identify properties of deep neural networks (DNNs) that confer generalization and robustness to distribution shift. I believe that understanding the information-theoretic limitations of Deep Learning is a key step to this end, and may help establish predictive performance guarantees for safety-critical settings.
My research interests are motivated by the “adversarial examples” phenomenon, whereby artificial intelligence (AI) models may be fooled by input manipulations that are deemed by many humans to be imperceptible or irrelevant. Adversarial examples highlight confirmation biases associated with common characterizations of representation learning in DNNs. For example the contribution of visually salient natural image features to categorical predictions may be over-estimated.
As part of my broader interest in trustworthy AI systems, I believe that more systematic ethical oversight of AI research is required to maintain trust in the field, and to ensure that respect for (often digitized) persons remains at the forefront. Beyond my primary research objectives, I maintain a keen passion for applied AI projects that have the potential for positive social and/or environmental impacts.
news
Jul 19, 2022 | Submitted Bounding generalization error with input compression: An empirical study with infinite-width networks to Transactions on Machine Learning Research. A great collaboration between the Guelph MLRG and Anna Golubeva (MIT/IAIFI), Mihai Nica (Guelph/Math & Stats), and Yani Ioannou (UCalgary). We show how to lessen the need for a train/validation/test split by training models to compress the training set, strengthening the connection between mutual information and DNN generalization. |
Jun 27, 2022 | New workshop paper led by Mohammed Adnan Monitoring Shortcut Learning using Mutual Information accepted at ICML 2022 Workshop on Spurious Correlations, Invariance, and Stability. |
Mar 7, 2022 | New journal article Predicting dreissenid mussel abundance in nearshore waters using underwater imagery and deep learning resulting from my collaboration with Environment and Climate Change Canada (ECCC) was published in Limnology and Oceanography: Methods. I presented the work at the SOLE’22 conference, and it was featured in UGuelph News as well as Great Lakes Now. |