I am a principal research scientist/director at Google. These days, I analyze complex machine learning models. I have also worked on question-answering systems, ad auctions, security protocol analysis, privacy, and computational biology.

See my resume and publication history for more details.


To track updates to this site:

Work on Attribution

TrackIn:

    • Answers the question: How important is a training data point to a model's prediction?
    • The techniques works for any model trained on a variant of Stochastic Gradient Descent (including Adam, Adagrad etc.)
    • Paper
    • FAQ

Integrated Gradients

    • Answers the question: How important is a feature to a model's prediction? The feature could be a pixel of an image, a word from a sentence, an atom in a molecule etc. The prediction could be the probability of cancer, a sentiment score, or a effectiveness of a molecule as a drug.
    • The main technical idea is to use path integrals of the gradients of the output with respect to the features.
    • Paper (Appeared at ICML 17)
    • GitHub code
    • Using Integrated Gradients to construct attacks on models
      • Paper (appeared at ACL18)
    • Principled visualizations of feature attributions:

Conductance

    • Answers the question: How important is a neuron to a deep network's prediction?
    • The technical highlight is a composition of Integrated Gradients and Backpropagation (Chain Rule).
    • Paper (Appeared at ICLR19)

Shapley-Taylor Values

    • Answers the question: How important is the interaction between a pair or set of features to a model's prediction?
    • The technical highlight is a connection between the Taylor Series and the Shapley value.
    • Paper

Other Papers on Deep Learning Explanations

    • There are many Shapley values
      • Discusses how different applications of the Shapley value yield different attributions (feature importance scores)
      • Paper
    • A critique about a critique of our work.

Axiomatic Attribution for Multi-linear Functions

    • Discusses how to attribute the change in the output of a funnel to the various stages in the funnel. This is useful to analyze changes in business metrics.
    • Paper
    • The algorithm is used within the AdWords Frontend to help Google's advertisers.

The Cascading Analysts Algorithm

    • Discusses how to summarize the change in a business metric along the product of hierarchical dimensions.
    • Paper
    • The algorithm is used within the AdWords and Google Analytics.

Trade-offs in Cost-sharing mechanisms