Megha Srivastava

Hi! I am a PhD student in the Computer Science department at Stanford University. I am co-advised by Dorsa Sadigh and Dan Boneh and study various topics within machine learning, including:

Methods to make ML models more robust, including mitigating bias amplification in online systems [1] and scaling verification of third-party model training services [2].

Natural Language as a powerful interface to improve model reliability, such as incorporating humans' beliefs of the underlying causal model [3], informing shared latent actions for robotic control [4], and constraining policy learning to be more interpretable [5].

Applications in education, ranging from second-language learning [6] to motor control skills [7, 8]. How can we model human learning dynamics over time, help agents effectively share skills with each other, and automatically design training curricula in specialized tasks?

Evaluation beyond static benchmarks. For example, programmers can introduce security vulnerabilities when relying on code-generation models [9], and we see similar patterns of misplaced confidence when language models are used for information-seeking tasks [10, 11]. What are the diverse ways humans adapt their behaviors when interacting with AI systems?

I received my B.S. in Computer Science with a Creative Writing minor from Stanford, working with Tatsu Hashimoto and Percy Liang on distribution shifts and fairness in machine learning, and Kalanit Grill-Spector on generalization in visual perception. I have also spent time in Cambridge, MA visiting MIT LINGO, was an AI Resident at Microsoft Research Redmond, and worked with Hoda Heidari and Andreas Krause through the amazing ETH Zürich Summer Research Fellowship.