[an error occurred while processing this directive] Shibani Santurkar [an error occurred while processing this directive] [an error occurred while processing this directive]
[an error occurred while processing this directive] Shibani Santurkar [an error occurred while processing this directive]
[an error occurred while processing this directive] Shibani Santurkar [an error occurred while processing this directive]
[an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] PhD Candidate [an error occurred while processing this directive] Massachusetts Institute of Technology [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive]
  • Artificial Intelligence
  • [an error occurred while processing this directive] From ImageNet to Image Classification: Contextualizing Progress on Benchmarks [an error occurred while processing this directive] Large-scale benchmarks have been instrumental in guiding recent progress in machine learning and are typically regarded as the ground truth during model development. But is this de-facto treatment of benchmarks as the gold standard truly justified?

    Through a case study on the popular ImageNet dataset, we demonstrate an inherent tension between the goal of building challenging, realistic datasets and leveraging scalable dataset collection pipelines to do so. Specifically, we study how design choices in the ImageNet creation process impact the fidelity of the resulting dataset---including the introduction of biases that state-of-the-art models exploit. Our analysis pinpoints how a noisy data collection pipeline can thus lead to a systematic misalignment between the resulting benchmark and the real-world task it serves as a proxy for.

    Joint work with Logan Engstrom, Andrew Ilyas, Aleksander Madry and Dimitris Tsipras. [an error occurred while processing this directive] Shibani Santurkar is a PhD student in the MIT EECS Department, advised by Aleksander Mądry and Nir Shavit. Her research revolves around two broad themes: developing a precise understanding of widely-used deep learning techniques; and identifying avenues to make machine learning be robust and reliable. Prior to joining MIT, she received a bachelors degree in electrical engineering from IIT Bombay, India. She is a recipient of the Google Fellowship. [an error occurred while processing this directive] Personal home page [an error occurred while processing this directive] [an error occurred while processing this directive]