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 human-AI interaction, including:

Methods for robust machine learning, including mitigating bias amplification in online systems [1] and controlling GPU nondeterminism to enable verifiable training [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].

Education applications, ranging from second-language learning [6] to embodied control skills [7, 8]. How can AI help us model student learning dynamics over time, identify the skills different individuals struggle to learn, and provide assistance that improves learning?

Interactive evaluations that extend 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, like question answering and solving crossword puzzles [10]!

I previously studied Computer Science and Creative Writing at 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 and neuroscience. 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 on algorithmic fairness through the amazing ETH Zürich Summer Research Fellowship.