Rosemary is a third-year PhD student at Polytechnique Montreal and the Mila institute. Her primary research interests center around developing novel machine learning algorithms inspired by human learning and abilities.

Many real world applications such as speech recognition, language modeling and image captioning are sequence processing tasks and they are commonly trained using recurrent neural networks (RNNs). One difficulty with RNN training is to generate more consistent samples and plan for the longer-term future. Some of her work has been focused on utilizing variational methods and auxiliary losses to help overcome these issues for supervised, unsupervised, and model-based reinforcement learning tasks. Another issue with RNN training is correct credit assignment through time. She has been working on alternative ways of training RNNs that can correctly assign credit while retaining a biologically plausible training procedure.

While developing novel algorithms, she has noted the importance of ensuring machine learning research is reproducible and fair so that the impact of these novel algorithms is positive for all members of society. She has been involved in organizing the ICLR reproducibility challenge and workshops to help with these issues.

For links to papers, open-sourced codes and videos, please visit her website