Ipsita Ghosh

ML Researcher • Optimization • Low-Rank Learning

Verified record

Publications & Patents

This page highlights the work I most want people to read first: efficient low-rank training, geometry reconstruction, and applied optimization work that made it into patented systems.

Entries below were checked against conference records, OpenReview, and patent databases.

Research papers

Featured publications

NeurIPS 2025 Poster

Q3R: Quadratic Reweighted Rank Regularizer for Effective Low-Rank Training

Ipsita Ghosh, Ethan Nguyen, Christian Kümmerle

Introduces a low-rank regularization method for training low-rank models directly, instead of relying only on low-rank fine-tuning after dense pretraining.

  • Frames Q3R through an IRLS-inspired objective for low-rank optimization.
  • Reports strong compression results on Transformer-style models, including ViT-Tiny.
  • Positions the work around practical low-rank pretraining rather than only parameter-efficient adaptation.
NeurIPS 2024 Poster

Sample-Efficient Geometry Reconstruction from Euclidean Distances using Non-Convex Optimization

Ipsita Ghosh, Abiy Tasissa, Christian Kümmerle

Studies how to recover geometric structure from sparse pairwise distances using non-convex rank minimization and an IRLS-style algorithm.

  • Gives a local convergence guarantee under minimal random distance sampling assumptions.
  • Establishes a restricted isometry property tailored to the tangent space of low-rank symmetric matrices.
  • Shows stronger data efficiency than prior baselines on simulated and real-world geometry reconstruction tasks.

Intellectual property

Patents from industry work

Inventor order below follows the patent records for each individual grant and related publication.

Granted patent May 2024

US11994936B2: Automated optimization of error-handling flows in memory devices

Jay Sarkar, Ipsita Ghosh, Vamsi Pavan Rayaprolu

Covers automated reordering of SSD error-handling operations using latency and recovery information, with the goal of improving memory-system performance under real workloads.

Granted patent June 2024

US12019874B2: Adaptive optimization of error-handling flows in memory devices

Jay Sarkar, Vamsi Pavan Rayaprolu, Ipsita Ghosh

Extends workload-aware optimization of memory-device recovery flows, emphasizing adaptive selection of error-handling behavior under changing operating conditions.