“While existing neural network models perform surprisingly well on several linguistic tasks, there is growing evidence that these models rely on non-human like heuristics and strategies. For example, even participants on MTurk, who are not experts in linguistics or machine learning, can quite easily come up with examples that can fool these models. We plan to analyze thousands of such non-expert generated examples that fooled state-of-the-art models in order to gain insight into some of the heuristics that these models rely on.
I am very excited about this project because I will get an opportunity to apply some of my existing skills to analyze forms of data that I’ve never worked with before in a domain I am not very familiar with!
I am also really looking forward to getting more exposure to several aspects of working in the industry, like working in a team, for example. I have found that when I am working on a certain project (or a set of related projects) for a long period of time, it can feel like I am stuck in a bubble. I am grateful for this opportunity to break out of this bubble a bit and explore new ways of thinking and working.”