Research

Selected Publications

• 

  Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding

  Lakshay Sharma, Alex Marin

  arXiv 2026

  We use reverse-geocoded text alongside semantic IDs as self-supervised training signals so that location encoders learn administrative boundaries and place names. This improves existing encoder architectures, and our RevGeoRQ architecture outperforms contemporary location encoders on several downstream tasks. The human-readable reverse-geocoded strings and discrete semantic IDs also make the learned representations more interpretable.

• 

  Subimage Overlap Prediction: Task-Aligned Self-Supervised Pretraining For Semantic Segmentation In Remote Sensing Imagery

  Lakshay Sharma, Alex Marin

  WACV 2026 Workshop: Computer Vision For Earth Observation (CV4EO)

  We propose a self-supervised pretraining task for remote sensing imagery: predicting where a sub-image is located within its source image. It speeds up convergence and matches or beats other SSL methods on semantic segmentation while using far less pretraining data.

• 

  Landsat-Bench: Datasets and Benchmarks for Landsat Foundation Models

  Isaac Corley, Lakshay Sharma, Ruth Crasto

  ICML 2025 Workshop: TerraBytes - Towards global datasets and models for Earth Observation

  We introduce three benchmarks for Landsat foundation models, adapted from EuroSAT, BigEarthNet, and LC100, and show that models pretrained on Landsat imagery outperform ImageNet-pretrained ones on downstream tasks.

• SuoiAI: Building a Dataset for Aquatic Invertebrates in Vietnam

  Tue Vo, Lakshay Sharma, Tuan Dinh, Khuong Dinh, Trang Nguyen, Trung Phan, Minh Do, Duong Vu

  ICLR 2025 Workshop: Tackling Climate Change with Machine Learning

  We propose an end-to-end pipeline for building a dataset of aquatic invertebrates in Vietnam and classifying species with ML, using semi-supervised learning to reduce annotation effort.

• Natural Language Understanding with the Quora Question Pairs Dataset

  Lakshay Sharma*, Laura Graesser*, Nikita Nangia*, Utku Evci* (*equal contribution)

  arXiv 2019

  We compare linear, tree-based, and neural models for duplicate question detection on the Quora dataset, finding that a simple Continuous Bag of Words model outperforms more complex recurrent and attention-based models.

Patents

• Temporal Localization of Mature Content in Long-Form Videos Using Only Video-Level Labels

  Lakshay Sharma et al.

  US Patent 11829413

• Generating Improved Traffic Speed Data for Road Segments in a Geographical Area Using Traffic Speed Prediction Neural Networks

  Lakshay Sharma et al.

  Patent pending (United States)

A full list of publications can be found on my Google Scholar page.