Rohan Wadhawan

Empirical analysis of pruning strategies in Federated Learning

• Investigated the effect of pruning on model generalization and accuracy vs. efficiency tradeoff in federated setup.
• Proposed protocol applied to the Iterative Magnitude Pruning method achieved an improvement of 6% on FMNIST and 8% on MNIST for a 99% prune ratio when compared with one iteration of standard federated training.
• Skills used: Python, PyTorch, Google Colab, Latex, Overleaf

Spotify music recommendation using Reinforcement Learning

• Simulated Spotify music recommendations, precisely predicting skips in Spotify streaming data as a Reinforcement Learning Problem.
• Employed Deep Deterministic Policy Gradient framework and Offline Reinforcement learning techniques to recommend a diverse list of songs that reduced the average skip rate by 12% from the baseline. Go to Project Page
• Skills used: Python, TensorFlow, Google Colab, Git

Face Synthesis using Descriptions Extracted from Unstructured Text

• Invented a novel pipeline to generate faces from their corresponding textual description. The motivation was to augment the reading experience for young children, especially those with reading difficulty, by animating characters through facial cues. Go to Publication
• Developed a crowdsourcing platform and consolidated our Multi-Attributed and Structured Text-to-face (MAST) dataset consisting of structured textual descriptions for face images.
• Performed text classification to filter out descriptive sentences from a textual data consolidation of Gutenberg and Face2Text datasets using Bi-LSTM with attention mechanism and achieved 98.5% accuracy and 0.97 F1 score on the test set
• Devised an algorithm for fast transformation of an unstructured facial description to a structured one; it has linear complexity with respect to the number of words in the sentence
• Trained an Attentional Generative Adversarial Network to synthesize faces from structured descriptions and reported benchmark scores of 54.09 Freechet' s Inception Distance, 1.080 Facial Semantic Distance, and 60.42% Facial Semantic Similarity on our MAST dataset
• Skills used: Keras, TensorFlow, PyTorch, OpenCV, Google Colab, Microsoft Cognitive Service, Angular Framework, MongoDB, NodeJS, Heroku, Overleaf

Project ViSTARa - India Winner UN Reboot the Earth Hackathon

• Designed a web-based Learning Management platform to Educate rural women through India's Self-Help Group network and Empower them to lead green climate initiatives. Coverage
• Devised a recommender system to suggest crops that require minimum supplementary irrigation based on rainfall patterns and Employed predictive analytics to detect deforestation at the district level in India

GRiD Flipkart Machine learning challenge for Large Scale Object Localization

• Architected a ResNet-34 inspired model to perform object localization on Flipkart’s large and diverse items dataset.
• Trained the model on images of size 128x96 (downscaled from VGA to a 1MP camera resolution)
• Achieved an IoU score of 90.05% on the private test set.
• Skills used: Keras, TensorFlow, OpenCV, Google Colab.

Skillset recommendation system to aid engineering aspirants in securing an Internship

• Proposed a skill set recommender system to aid engineering aspirants in securing an Internship. Go to Project Page
• Consolidated a small dataset of various skills an aspiring intern may have and categorized them as generic, company-specific, and domain-specific.
• Modeled the skill selection problem as a combinatorial optimization problem with multiple objectives and employed a Genetic algorithm (GA) to solve it.
• Established a Fitness Function to evaluate the fitness of each individual in the chromosome population.
• Formulated an Objective Function to combine opposing goals of finding the best skillset while minimizing the time to achieve it.
• Implemented a modular GA pipeline in C++ to evaluate and select the optimum set from the possible combinations of GA operations: population initialization, parent selection, crossover, mutation, survivor selection, and termination.
• Skills used: C++.

Book My Flight - Database Management Semester Project

• Implemented a Flight Booking Management system for domestic flights in India.
• Modeled a MySQL database system with a complex database trigger and recovery mechanism.
• Designed Java-based user interface. Go to Project Page
• Skills used: Java, MySQL, ERDPlus.

• Introducing ConTextual: How well can your Multimodal model jointly reason over text and image in text-rich scenes?

  Go to HuggingFace Article

• Synopsis: Multi-Attributed and Structured Text-to-Face Synthesis

  Go to Medium Article

• Synopsis: Intelligent Monitoring of Stress Induced by Water Deficiency in Plants Using Deep Learning

  Go to Medium Article

• VQ-GAN & Transformer — Taming Transformers for High-Resolution Image Synthesis: Synopsis

  Go to Medium Article

• Reboot the Earth: Project ViSTARa aims to educate women in rural areas on climate changes

  Go to Article on First Post