Nikhil Mehta

I am a Staff Research Scientist at Google DeepMind, where I lead research on GenRetrieval and Semantic IDs for personalization in Large Language Models.

I received my Ph.D. from Duke University, advised by Professor Lawrence Carin, with a thesis on Continual Learning of Deep Neural Networks. During my PhD, I interned at Meta FAIR, Google Brain, and Google AutoML. Prior to Duke, I was a research associate at IIT Kanpur, where I worked with Professor Piyush Rai and Professor Gaurav Pandey.

Email  /  Google Scholar  /  Linkedin /  Github

My research focuses on Generative Retrieval, personalization with LLMs, and deep learning. For a complete list, see my Google Scholar.

Research on video tokenization recipes for understanding and VQA tasks. Preparing and implementing evaluation benchmarks for VQA using YouTube videos.

Adapting pre-trained language models for industrial-scale generative recommendations at Google.

A framework for holistic evaluation of recommender systems powered by generative models.

Semantic IDs for representing items in corpus, enabling better generalization in ranking for recommendations.

Methods for aligning large language models with recommendation knowledge for improved personalization.

A novel method for out-of-distribution detection and classification in few-shot learning settings.

A meta-learned approach for attribute self-interaction in continual and generalized zero-shot learning.

Invented GenRetrieval for recommendation systems, now adopted across multiple companies including Spotify, Meta, Snap, Amazon, Kuaishou, and Doordash.

Methods for pushing the efficiency limit of deep networks using structured sparse convolutions.

A task-specific feature map transformation strategy for continual learning, which adds minimal parameters to the base architecture while outerperformaing previous methods in discriminative and generative tasks.

A Bayesian nonparametric framework using shared rank-1 weight factors for federated learning.

Continual Learning using Bayesian Nonparametrics and layer-wise dictionary of weight factors.

Counterfactual Representation Learning with Balancing Weights

A Gamma Ladder Variational Autoencoder for graph-structured data that brings together the interpretability in hierarchical, multilayer latent variable models and the strong representational power of graph encoders.

A deep generative framework for overlapping community discovery and link prediction combining the interpretability of stochastic blockmodels, such as the latent feature relational model, with the modeling power of deep generative models.

We propose using Stein's Method to estimate the parameters of a distribution based on given samples. We also estimate the score function of the empirical distribution and propose a new generative model combining Stein score matching with the Langevin flow.

We present a probabilistic framework for multi-label learning based on a deep generative model for the binary label vector associated with each observation.

A Bayesian nonparametric model for jointly inferring the time-to-event and the individualized risk-based cluster assignments.

A mutual information based algorithm for the estimation of full 6-degree-of-freedom (DOF) rigid body transformation between two overlapping point clouds.

Forked from Jon Barron's page.