Nitin Kishore Sai Samala

Nitin Kishore Sai Samala

Staff Data Scientist/ML Research Engineer | Deep Reinforcement Learning | UMass Amherst CS Grad | NLP | BITSian | Research Enthusiast | Quick Learner

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

Posts

Green Screen Image Composition Transfer

less than 1 minute read

Published:

Using StableDiffusion for text2img background generation and MODNET/U2net model to seamlessly superimpose the isolated foreground using Alpha-matting and apply Nvidia’s FastPhotoStyle for an image stylization and smoothening

DeepfakeCapsuleGAN

less than 1 minute read

Published:

Using GANS to generate images that are then used for deep fakes. Using Capsule networks instead of regular CNNs

Improving Open Domain Dialogue-Systems (Chatbots)

less than 1 minute read

Published:

Built a seq2seq neural conversational model in PyTorch using attention with intention and a diversity promoting objective function to prevent irrelevant generic outputs’

DSR Reinforcement Learning to navigate a Labyrinth

less than 1 minute read

Published:

Generalized Successor Representations from Neuroscience within an end-to-end deep reinforcement learning framework, comparing its efficacy to DQN on two diverse environments (Mazebase and DOOM) given raw pixel observations.’

portfolio

publications

Exploring the Efficiency of Capsule Networks in GANs

Published in Umass Amherst, 2018

Used Capsule Network as discriminator for a generative adversarial network (GAN), trained using several hacks, which outperformed CNN-GANs at modeling image data distribution of mnist, cifar10 and celebA

Recommended citation:

A Multi-Objective Optimization for Clearance in Walmart Brick-and-Mortar Stores

Published in INFORMS Journal on Applied Analytics, 2021

Using RL to simulate price markdowns in stores, personalized to region and demographic based optimal policy in order to balance contradictory multi-objective optimization scenario. We need to reduce the price to increase sell through rate of bad performing products, but this decay has to be done in a slow controlled duration, so as to avoid triggering a replenishment before shelf change cycle. This involves reward design with multiple cost penalties. For single price change scenario we devised a variant of the Black Scholes Equation

Recommended citation: Nitin Kishore Sai Samala, Yixian Chen, Prakhar Mehrotra, Kamilia Ahmadi, Viresh Jivane, Linsey Pang (2021) A Multiobjective Optimization for Clearance in Walmart Brick-and-Mortar Stores. INFORMS Journal on Applied Analytics 51(1):76-89. https://doi.org/10.1287/inte.2020.1065 https://doi.org/10.1287/inte.2020.1065

InAppropriate Content Detection System

Published in US Patents, 2021

Built an end-to-end automated pipeline for “Inappropriate content detection” for UGC(user generated content) items on Walmart eCommerce using Vader Sentiment, Topic Modelling, Snorkel and BERT. Built as part of the Trust and Safety framework. This framework creates auto-generated keyword based rules that are obtained through Topic Modelling, over a multi-processed scraper that pulls in RSS feeds of news articles and summaries

talks

teaching