Ananth Balashankar
Senior Research Scientist, Google DeepMind
New York
ananth(at)nyu(dot)edu
[CV]   [Research Statement]   [LinkedIn]  

I'm a Senior Research Scientist at Google DeepMind in the GenAI Safety and Alignment team in New York. I have contributed to the post-training of Gemini 2.0, 2.5, AI Overviews and AI Mode. Specifically, my research interests lie in adversarial training, reward model alignment and robustness. You can find some of our recent papers here.

Previously, in Google Research, I worked on improving the robustness and safety of classifiers and foundation models (LLMs) that power applications such as YouTube ads monetization, Bard, and Search Generative Experience. I led work from basic research to product impact, with >10 launches that reach 1B+ users.

I got my Ph.D in Computer Science advised by Prof. Lakshminarayanan Subramanian at NYU's Courant Institute of Mathematical Sciences and Dr. Alex Beutel (now at OpenAI). I was a Student Researcher at Google AI (2019-22), where I worked on counterfactual text robustness. I was a Software Engineer at Google for 3 years, where I worked on recommendations at Google Play Store and the Play Developer Console in Mountain View and London. I graduated from the Indian Institute of Technology, Kharagpur with B.Tech/M.Tech in Computer Science advised by Prof. Niloy Ganguly.

Honors

  1. NYU Janet Fabri Prize for best Ph.D dissertation in Computer Science - 2023
  2. Google Student Research Advisor Program Fellowship (2019-2022)
  3. Spot bonus for research contributions in the Google Responsible AI team - 2021
  4. NYU Harold Grad Memorial Prize for promising Ph.D achievement - 2019
  5. Best Paper Award at NAACL 2024 TrustNLP, and ICML 2019 AI for social good workshops
  6. MacCracken Fellowship (2017-22)

Research Interests

I am interested in building safe and responsible ML models through methodologies including domain faithful optimization, data augmentation and causal feature selection. Broadly, my work has had demonstrable business and research impact across five real world application domains:
  1. Safety and Responsibility in AI
    Automated detection of online toxic comments improves the quality of interaction in social media. However, the variations in context of comments make it hard to protect specific demographic groups from disparate impact. By explicitly modeling such nuances through counterfactual data augmentation, we improved the accuracy of detecting toxicity by 6% Through this publication at EMNLP '21, a premiere NLP conference, I have fostered deep engagements with Google's Responsible ML team. I have also deployed ML models that optimize business objectives like diversity at Google Play.

  2. ML Robustness
    Robust ML models are critical in building high-stake applications and require a shift from traditional ML models that focus only on optimizing accuracy over the observed but limited test data. By incorporating rules and data from the real world, we have improved accuracy of state-of-the-art transformer based models by 12% in this publication at WSDM '21, the premiere data mining conference.

  3. Causal-Aware ML
    Causality based question answering lies at the core of customer support tools like chatbots. Prior ML models fail to capture the directed nature of causality, for example rain causes traffic delay, and not vice versa. By learning asymmetric causal embeddings faithful to causal graphs, we improved accuracy on Yahoo! Answers by 21% in this paper at ACL '21, a premiere NLP conference.

  4. AI for Social Good
    Forecasting famine is critical for the mobilization of aid to millions of people, but hard to solve due to data scarcity in fragile countries. By building a news-based causal-aware forecasting framework that extracts causal features from 11.2 million news articles across 2 decades in 21 fragile countries, we have improved forecasting accuracy by 32% compared to state-of-the-art predictive models. This paper is accepted at IC2S2 '21, the premiere computational social science conference, and at Science Advances in 2023. The tool will be used by the World Bank Data Science group for aid allocation on food security. Based on this research, a few co-authors have founded a socio-economic inference start-up Velai, Inc.

  5. ML in Privacy
    Corporate privacy compliance policies are legally prescriptive, but not directly enforceable in computer systems. By using the theory of contextual integrity through post-processing mappings, we have improved the accuracy of BERT-based deep learning models by 6% to extract privacy parameters for SQL-based enforcement in this paper at WWW' 19, the premiere web research conference.

Select Publications

Adversarial Safety

Reward Alignment

Robustness

AI for social good

AI for privacy

Other Research

Teaching Experience