Why do you care about AI Existential Safety?
While we chase the next frontier of machine learning performance as researchers, we are increasingly impacted by the potentially wide adoption of machine learning algorithms in many aspects of our daily lives. Prediction errors made by machine learning algorithms may amount to a 1% drop in the performance reported in a research paper; but when the algorithms are deployed, this 1% drop will have profound implications for the lives of the 1% (and often much more) of individuals affected by it. To this end, I care about the reliability and safety of machine learning algorithms and the alignment of machine learning algorithms with the tasks they are deployed in.
Please give at least one example of your research interests related to AI existential safety:
I study the reliability of machine learning algorithms from causal and probabilistic perspectives. From the causal perspective, I study what leads machine learning models to make certain predictions. Understanding the causality underlying the data-generating process and the algorithm's decision-making process could help us align the algorithms better with the actual tasks they are deployed in. From the probabilistic perspective, I am interested in building algorithms that know what they don't know. When the data cannot assist us with our questions of interest, the algorithms ideally avoid random guessing but acknowledge its ignorance.