I received my PhD from the Machine Learning Department and the Center for the Neural Basis of Cognition at Carnegie Mellon University. I was a member of the Brain Image Analyis Group led by Tom Mitchell.
- Speaking at the MAIN 2019 conference (Montréal Artifical Intelligence and Neuroscience).
- Co-organizing a workshop titled Minds vs. Machines: How far are we from the common sense of a toddler? at CVPR 2020.
- Our papers on using brain activity to interpret NLP models, to add brain-relevant bias to BERT, and to study the similarity of representations from different computer vision tasks have been accepted at NeurIPS 2019.
- Co-organizing a workshop on Context and Compositionality in Biological and Artificial Neural Systems at NeurIPS 2019.
- Received the 2018 Google Faculty Research Award.
- Spoke at the 2019 workshop on Semantic Processing and Semantic Knowledge at Dartmouth.
- Spoke at the Machine Learning in Neuroscience: Fundamentals and Possibilities SfN online conference.
- Spoke at the 2019 Cognitive Modeling and Computational Linguistics NAACL workshop.
- Spoke at the 2018 Workshop on Analyzing and interpreting neural networks for NLP.
- Gave a full day tutorial at the 2018 Society for Psychophysiological Research Annual Meeting.
- Was interviewed for the book Artificial Intelligence: Teaching Machines to Think Like People.
I use functional Magnetic Resonance Imaging (fMRI) and Magnetoencephalography (MEG) to investigate how the brain represents the meaning of words, sentences and stories.
FMRI and MEG record brain activity but yield very high dimensional, noisy images that are expensive to acquire. The number of data points in a typical experiment is therefore many orders of magnitude smaller than the number of data dimensions. Furthermore, there is a considerable subject-to-subject variability of brain anatomy. Combining data from multiple subjects is consequently a hard problem. Part of my work is finding machine learning solutions to these problems.
Another part of my work is defined by the complexity of language and the non-existence of a comprehensive model of meaning composition: we do not know how the meaning of successive words combine to form the meaning of a sentence. Investigating the brain representation of a sentence is therefore a complex task because we are both looking for the neural signature and trying to approximate the composition function. However, with appropriate experimental design and computational models, we can study both problems: we can use existing models of language to study the brain representation of meaning, and we can use brain data to evaluate different meaning composition hypotheses.