About Me
Hi, I’m Jaydeep! I’m obsessed with lot of things, especially understanding the complex ML algorithms in the most intuitive manner, interpretability of Deep Learning systems, solving some real world problems using AI that can make the world a better place to live. My long term goal is to create ML models which are robust, could able to learn from very few samples, generalize beyond iid settings and can do a high level reasoning about the world. I consider myself polymath and I love connecting dots between AI and different fields.
I’m currently pursuing Masters in Computer science from Indiana University Bloomington. Prior to joining IUB, I was working as a Assocaite Data Scientist for a Indian start up called : ‘ParallelDots’.
If you are also excited about any of the topics that I mentioned above and want to share your thought, feel free to reach out to me. You can mail me or reach out to me via LinkedIn.
Thanks for being here.
Projects
This is one of the coolest application of computer vision!
This project is based on an implementation of NeRF(https://www.matthewtancik.com/nerf) paper to render a novel 3D view of a scene. The objective of this project was to learn about the 3D view reconstruction, neural rendering and depth estimation using latest state of the art model in computer vision.
Supervised Contrastive Learning for pretrained language models
github.com/Jd8111997/Supervised-contrastive-learningSupervised contrastive learning proves a great technique for few shot learning!
This project is based on an implementation of supervised contrastive learning for pretrained language models paper(https://arxiv.org/abs/2011.01403).The main objective behind this project was to learn about few shot learning for data hungry large language models. We got fascinating results to see that after applying supervised contrastive loss; the model is able to learn a very good representation of discriminative features from very few samples.
I built it to enhance my crappy guitar riffs!
This purpose of working on this project was to explore the domain of speech enhancement by incorporating Speech Enhancement Generative Adversarial Networks.
Feeling bored, let's make an AI model to generate some dog images:)
This project is based on an implementation of Autoregressive Generative Models such as PixelCNN, PixelCNN++ and GatedPixelCNN to generate a novel images of a dog.
After watching Silicon Valley TV series, I got inspired to work on decentralized chatting application.
The main objective for working on this project was to learn more about decentralization concepts and algorithms by developing a decentralized chat application based on popular open-source matrix protocol and synapse server using Angular as front-end and Python-twisted as a backend.
Isn't it too tiring to manually delete duplicate images from computer ?
This project was a PyQt based GUI application that can detect Duplicate Images by sensitivity hasing algorithm; It can also detect the noisy images and can automatically delete them based on the threshold.
Collaborative filtering won't work if you don't have enough data.
Beats is a music hosting site and music recommendation engine built in C# using collaborative filtering algorithms.
Experience
Do you know, Failure to ensure accurate tracking of inventory can cost millions of dollars every year even within a single warehouse!!!
- Image Quality Assessment(IQA): Developed a new ’No reference Image Quality Assessment’ method using deep learning that can help to assess an image quality and helps to locate the distorted regions in an image in a multi task setting.
- Data valuation: Worked on data valuation methods to quantify the sample quality using meta reinforcement learning approach.
- Image enhancement: Worked on developing a pipeline for Image enhancement using a state of the art image debluring methods.
ParallelDots Inc.
Associate Data Scientist - DL Research
Nov 2020 - Jul 2021
https://www.paralleldots.com/
Solving some amazing problems in shelf retail execution using AI.
- Object Detection Benchmarking: Contributed in benchmarking of Domain Invariant Object Detectors across various client datasets and tried various state of the art object detection models to replace generalized object detector in ShelfWatch(main product of the company).
- Training Object Detectors & Classifiers: Training object detection models and product classifier on various client datasets to achieve high accuracy and good generalization on out of distribution classes.
- Object Detection for Mobile Device: Training a object detection models for mobile devices.
- Research Work: Contributed in a ongoing research work in semi supervised learning for dense object detection and our work got published in CVPR 2021 retailvision workshop.
Bhaskaracharya Institute For Space Applications and Geo-Informatics.
Software Engineering Intern
Dec 2018 - Mar 2019
https://bisag-n.gov.in/
Credits goes to a Pied-piper's idea of new internet!!
- I Developed a decentralized chatting application for BISAG scientists based on open source framework : ’Matrix Synapse ’, with functionalities such as user registration, emoji and multimedia file sending, birthday wish, delete a message, create and change group-name etc; using Python, Angular and SQLite.
Education
Indiana University
Masters in Computer Science
Aug 2021 - May 2023
IU is considered one of the most beautiful campus in the US.
- During my masters, I have studied many interesting subjects : Applied Algorithms, Elements of AI, Machine Learning, Computer Vision, Advance OS, Applied ML for Computational Linguistics, Music Data mining.
- Currently, I’m doing Masters Thesis on generalization under distributional shifts, disentanglement and causality in deep learning under the guidance of Prof. David Crandall.
Skills
- Languages : Python, R, C/C++, JavaScript, Java SE, SQL, Bash, PHP
- Libraries : PyTorch, Scikit, NLTK, Matplotlib, TensorFlow, Keras, JAX, Pandas, Scipy, Plotly
- Tools : GIT, MySQL, SQLite, Docker, Wandb, Kubernetes, Latex
Honors and Awards
- Won a silver medal for Cassava Leaf Disease Classification challenge on kaggle(top 4% worldwide). - Jan 21
- Won a silver medal for SIIM-ISIC Melanoma Classification challenge on kaggle (ranked top 2% worldwide) - Aug 20
- Expert’ in Kaggle Competitions(Current rank 1174 out of 166,012)
Publication
- Semi-supervised Learning for Dense Object Detection in Retail Scenes: CVPR 2021 RetailVision Workshop
- Comparative study of GAN and VAE: International Journal of Computer Applications (0975 - 8887) Volume 182 - No.22, October 2018
A Little More About Me
Alongside tinkering with ML models, some of my other interests and hobbies are:
- Playing riffs on guitar(Currently I’m learning to play Nothing else matters by Metallica)
- Excercising
- Cooking
- Watching stand up comedies
- Reading fictions(Currently I’m finishing Brothers Karmazov)
- Musing and trying to connect dots between science, arts, philosophy and history
My Bucket list
- Build my own RV and travel across all the states in the US
- Visit a F1 circuit at Monte Carlo
- Learn surfing and sky diving
- Produce my own music