About Me
Hi, I’m Yifan!
I’m a postdoc with the EconCS group at Microsoft Research, New England. I obtained my PhD from Northwestern University, where I was advised by Prof. Jason Hartline. I received my B.S. in Computer Science from Turing Class at Peking University in 2020, where I was advised by Prof. Yuqing Kong.
I have a broad interest in the intersections between theoretical computer science, AI, and economics. Recently, my research focuses on AI trustworthiness.
My Research
My research develops the theory of proper scoring rules for understanding the trustworthiness of an AI system. In this framework, proper scoring rules benchmark and incentivize the rationality of components in the AI system.
Proper scoring rules are functions that assess the quality of a probabilistic prediction against the realized random variable. A scoring rule is proper if the expected score is maximized when the prediction matches the true distribution of the random variable. In game theory, a proper scoring rule incentivizes a rational agent to report their subjective true prediction. In statistical decision theory, proper scoring rules evaluate the decision payoff of a rational decision-maker when assisted by the prediction. In machine learning, proper scoring rules (a.k.a., proper losses) correctly identify the optimal predictor with the lowest expected loss.
In my research, I model complex components of an AI system as a rational agent optimizing for expected payoff, which applies to predictors, human decision-makers, and downstream decision problems that are generally unknown when an upstream AI system is deployed.
My research topics include:
- Optimal information elicitation (optimization of proper scoring rules), a theoretical foundation to my framework.
- Calibration, requiring predictive model outputs to be reliably interpretable as probabilities. I focus especially calibration error design.
- Obtaining provable guarantees from AI systems, such as truthfulness and no-regret.
- Benchmarks for human-computer interaction or human-AI interaction, theoretically justified framework for understanding human behavior in algorithm-assisted decision problems.