Mansi Mane

Mansi Mane photo

Email:mansimane(FIVE)(AT)gmail(DOT)com, CV , GitHub , LinkedIn

I am working as an Applied Scientist at Amazon AWS and my interests include Machine Learning applications in Computer Vision, and Natural Language Processing. I did masters from Carnegie Mellon University focused on Machine Learning.

Publications

Work Experience

  • Applied Scientist II, Amazon (AWS), Santa Clara, USA

    • Trained billion scale parameter NLP model from scratch with minimal loss in accuracy. To enable customers to train such models seamlessly on AWS, worked on SageMaker Model Parallel and HuggingFace integration which is being used by 30% of distributed training customers
    • Researched different batch size scaling algorithms which reduced training time by upto 3 times for ResNet training.
    • Developed deep learning approaches like Siamese and triplet network in TensorFlow by using multi-modal attributes like image and text data for complementary item recommendations.
    • Worked on making PyTorch available in deep learning containers which is being used by 10000 users per week.
    • Built deep learning infrastructure using the following AWS services: S3, EC2, ECR, SageMaker, CloudFormation, CloudWatch, CodeBuild, IAM.
  • Data Scientist, Walmart Labs, Sunnyvale, USA

    • Built machine learning models and data pipelines in Hive & PySpark for large scale item recommendations with 20 milion items.
    • Deployed matrix factorization model for personalized item recommendations for 10M users which resulted in 0.08% gain in add to cart rate in online A/B test
    • Developed siamese networks model for complementary item recommendations with in 0.1% gain in click through rate in online A/B test
    • Developed machine translation model to generate product titles for 20,000 items sold with voice assistants
  • Research Assistant, CyLab Biometrics Center, CMU, Pittsburgh, USA

    • Pre-processed data for face parsing using Fully Convolutional Instance Aware Semantic Segmentation.
    • Segregated medical images containing nucleus, cells and not containing those for pap-smear test by applying techniques like Gaussian smoothening, threshoding etc.
    • Synthesized 2D face images at different poses by modifying existing code to project 3D fitted morphable model for face into 2D.

Other Machine Learning Projects