Sumanth Nandamuri

Experience

• 

  SIP Lab IIT Kharagpur

  May 2018 - Jul 2018

  Deep Learning - Computer Vision Internship

  Task: Deep CNN for Semantic Segmentation of Thyroid from ultrasound images.◦ Implemented encoder-decoder fully convolutional network architectures SegNet and UNet in PyTorch.◦ Employed Transfeer learning by using weights of VGG-11 in encoder part of SegNet and UNet to improve dice coefficient from 0.7 to 0.8 .◦ Created a custom architecture SUMNet to further improve dice coefficient from 0.8 to 0.85 .◦ Used weighted loss function to give higher weights to the boundary pixels of the label to increase to accuracy at the boundary pixels. Show less

• 

  Microsoft

  May 2019 - Jul 2019

  Applied Scientist

  1. Multi Task Self Supervised Visual Representation Learning.2. Implemented existing and novel self supervised representation learning SSRL methods mask encoding, jigsaw, SimCLR and evaluated them on standard downstream tasks.3. Devised a novel quantitative method to (1) evaluate effectiveness of SSRL & (2) reason on building efficient Multi Task SSRL techniques.

• 

  Microsoft

  May 2020 - Dec 2023

  Machine Learning Engineer

  Bing Related Searches◦ Responsible for relevance of Bing Related Searches product, measured by gains in feature DAU, revenue, manual reformulations.◦ Deployed provider re-ranking to maximize relevance and RS engagement.◦ Transformer based Relevance classifier, generated training data using GPT4 teacher. Distilled XLM Roberta model to scale at web scale.◦ Deployed LightGBM based L2 ranker based on location and click intent.◦ Deployed transformer models (T5, GPT2, Phi1.5) to generate high click utility related suggestions.Bing Vision Team◦ Fine grained visual attribute recognition systems, training CNNs for 3000 class multi-label classification.◦ Handling weakly & noisy labelled data, class imbalance.◦ Implemented distributed training in pytorch from scratch.◦ Evaluation of model performance with human judges and improving models with active learning.◦ Highlights: ◦ Custom Loss funtions: MultiMaxLoss, FocalBCELoss for , AssymetricFocalBCELoss. ◦ Architectures: MultiHeadHierarchical CNN & ViT, momentum distillation, label smoothing. ◦ Distributed training and mixed precision training with basic PyTorch. ◦ Data: Sampling with replacement, count based sampling, RandAugment.Multi-modal pretraining.◦ Implemented multi-model pretraining for image-text representation learnnig.◦ Custom objective Masked Image Modelling (MIM) along with MLM and ITM objectives to improve vision recall.◦ Involved modifying transformer architecture, custom positional embeddings, loss formulations. Show less

• 

  Google

  Jan 2024 - now

  ML SWE

  Working on ads conversion modeling improvements