Ekansh Sharma

Curriculum Vitae

PDF version compiled from LaTeX alongside an accessible web summary.

Download the full formatted two-page PDF or browse the structured summary below.
Download PDF Open in New Tab

Your browser cannot display the embedded PDF inline.

Open CV PDF
Web Summary View (Mobile & Screen-Reader Friendly)

Research Profile & Contact

Ekansh Sharma — Machine Learning Engineer (Google (YouTube Shorts)) · Mountain View, CA
Email: () · Google Scholar · GitHub

Machine learning researcher and engineer specializing in large-scale sequential modeling, optimization geometry, low-precision training numerics, and statistical inference.

Professional & Research Experience

Google — Machine Learning Engineer — YouTube Shorts
September 2025 – Present · Mountain View, CA, USA
  • Generative Sequential User Modeling (TokenMinds): Co-authored and deployed an LLM-adapted encoder-decoder user model (370M MoE encoder / 370M dense decoder in JAX) unifying short-form, long-form, and search histories via hierarchical RQ-VAE Semantic IDs; serves billions of users asynchronously (1.44M read QPS, 96.4% cache hit rate), cutting upstream training and serving compute by 50% and 31% while driving statistically significant live A/B gains on YouTube Shorts (+0.13% Satisfied Engagement, +0.11% Engaged Users, +0.33% Freshness Engagement).
  • Long-Context Compression & Representation Stability: Extended sequential user modeling to long interaction horizons using soft-token compression, sparse cross-attention bottlenecks, and multi-resolution recency compression. Established longitudinal representation-stability diagnostics (cosine drift, perturbation invariance, split-history self-retrieval, and effective-rank monitoring) and non-autoregressive auxiliary reconstruction objectives to prevent representation collapse under continuous daily training.
  • Causal Point-Process Objectives, Variance Reduction & Cold-Start Estimation: Designed multi-task sequence objectives for noisy implicit-feedback streams, incorporating right-censored survival/hazard modeling, count-process supervision, deterministic importance/loss weighting, and C++ streaming input filters with strict request-time causality checks against lookahead label leakage; developed empirical-Bayes shrinkage representations for cold-start candidate scoring.
  • Optimization Symmetries & Low-Precision Training Numerics: Investigated orthogonal gauge symmetries and outlier-mitigation techniques for low-precision (FP8/INT4) quantization-aware training under matrix-sign and adaptive optimizers; profiled end-to-end XLA/TPU HBM, VMEM, and operator-fusion rooflines on large-scale multi-task ranking architectures.
  • Scaling Laws & Experimentation Infrastructure: Conducted Iso-FLOPS capacity, batch-size scaling, Maximal Update Parameterization (μP), and low-compute proxy transfer studies to predict full-scale ranking performance; led multi-launch model integration cycles and built automated training-lifecycle governance and paired statistical evaluation tooling for continuous TPU training pipelines.
Vector Institute & University of Toronto — Graduate Researcher (Adviser: Prof. Daniel M. Roy)
March 2018 – August 2025 · Toronto, ON, Canada
  • Loss-Landscape Geometry & Non-Local Model Merging: Established simultaneous linear mode connectivity modulo permutation symmetries across neural network populations (ECML-PKDD 2024; CPAL Spotlight 2024); identified deep-layer variance collapse in non-local multi-task model merging across distinct foundation models and open-sourced task-specific statistical repair (TACT).
  • Sparse Optimization Dynamics: Analyzed normalization-induced gradient skew in sparse training from scratch (SparseOpt, ICML 2026) and weight-symmetry permutation alignment for random-initialization lottery-ticket masks (ICML 2025).
  • Nonasymptotic Monte Carlo & Heavy-Tailed Random Graphs: Derived compositional nonasymptotic error bounds for nested MCMC in probabilistic programs and consistent bootstrap tail-index estimators for sparse exchangeable power-law graphs (Electronic Journal of Statistics, 2021).
Amazon.com — Applied Scientist Intern — Sponsored Display Ads (Manager: Ben Allison)
January 2021 – May 2021 · Edinburgh, UK
  • Counterfactual Evaluation & Contextual Bandits: Developed counterfactual off-policy evaluation estimators for ad-selection policies to quantify conversion lift while reducing live A/B exploration cost; evaluated Upper Confidence Bound (UCB) contextual bandit algorithms in production A/B experiments measuring click-conversion rate.

Education

  • Ph.D. in Computer Science, University of Toronto (Conferred November 2025) · Adviser: Prof. Daniel M. Roy
    Thesis: Exploiting linear connectivity modulo permutations in neural networks for parsimonious machine learning
  • M.Sc. in Computer Science, University of Toronto (Conferred 2018) · Adviser: Prof. Daniel M. Roy
    Thesis: Sparse exchangeable bipartite graphs
  • B.A.Sc. with Honors in Electrical Engineering, University of Toronto (Conferred 2016)
    Minor in Robotics and Mechatronics

Selected Publications & Preprints

  1. TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems. Qingyun Liu, Bo Yan, Yang Liu, Yuji Roh, Ekansh Sharma, Likang Yin, Emma Olowo, Yuxuan Li, Yuan Jiang, Diego Uribe, Saksham Aggarwal, Siqi Wu, Yuan Hao, Vikas Kedigehalli, Min-Hsuan Tsai, Lukasz Heldt, Lichan Hong, Li Wei, Xinyang Yi. Preprint arXiv:2606.25147 (2026).
  2. SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training. Mohammed Adnan, Rohan Jain, Tom Jacobs, Ekansh Sharma, Rahul G. Krishnan, Rebekka Burkholz, Yani Ioannou. International Conference on Machine Learning (ICML) (2026).
  3. Sparse Training from Random Initialization: Aligning Lottery Ticket Masks using Weight Symmetry. Mohammed Adnan, Rohan Jain, Ekansh Sharma, Rahul G. Krishnan, Yani Ioannou. International Conference on Machine Learning (ICML) (2025).
  4. The Non-Local Model Merging Problem: Permutation Symmetries and Variance Collapse. Ekansh Sharma, Daniel M. Roy, Gintare Karolina Dziugaite. Preprint arXiv:2410.12766 (2024).
  5. Simultaneous Linear Connectivity of Neural Networks Modulo Permutation. Ekansh Sharma, Devin Kwok, Tom Denton, Daniel M. Roy, David Rolnick, Gintare Karolina Dziugaite. ECML PKDD 2024 · CPAL 2024 Spotlight Track (2024).
  6. Bootstrap estimators for the tail-index and for the count statistics of graphex processes. Zacharie Naulet, Daniel M. Roy, Ekansh Sharma, Victor Veitch. Electronic Journal of Statistics, 15(1), 282–325 (2021).
  7. Approximations in Probabilistic Programs: A Compositional Nonasymptotic Analysis of Nested MCMC. Ekansh Sharma, Daniel M. Roy. International Conference on Probabilistic Programming (PROBPROG) (2020).
  8. Exchangeable modelling of relational data: checking sparsity, train-test splitting, and sparse exchangeable Poisson matrix factorization. Victor Veitch, Ekansh Sharma, Zacharie Naulet, Daniel M. Roy. Preprint arXiv:1712.02311 (2020).

Technical Skills

  • Languages & Systems: Python, C++, C, SQL, Bash; multi-slice TPU/GPU distributed training, custom C++ streaming data & sparse coordinate kernels, XLA/Pallas profiling, HBM/VMEM roofline analysis.
  • ML & Optimization: JAX/Flax, PyTorch, TensorFlow; LLM/Transformer sequence modeling, non-Euclidean/matrix-sign & sparse optimization, low-precision (FP8/INT4) training numerics, model merging, μP scaling.
  • Statistical & Quantitative Methods: Survival & point-process hazard modeling, counterfactual and off-policy evaluation, contextual bandits, empirical-Bayes shrinkage, heavy-tailed asymptotics, nonasymptotic concentration, MCMC, low-SNR paired/bootstrap inference.

Honors & Academic Service

  • Awards: Vector Research Grant (2018–2025); Department of Computer Science 50th Anniversary Graduate Scholarship (2017); University of Toronto Excellence Award (2013, 2015); Dean's Honors List (2011–2016).
  • Academic Service: Reviewer: NeurIPS (2021–2023, 2025–2026); Reviewer: ICML (2023–2025); Reviewer: ICLR (2020–2023).