Research
As a Graduate student at UMass Amherst, I work with Prof Marco Serafini and Prof Hui Guan on the topic of Machine Learning and Systems and I specialize in Graph Neural Networks. As an undergraduate student at IIT Bombay, I worked with professor Sanjay S. Pandey on Convolutional Neural Networks.
Graph Neural Networks
In collaboration with Prof. Marco Serafini and Prof. Hui Guan, I embarked on a comprehensive research journey exploring the dynamics of training Graph Neural Networks (GNNs). Our study involved an extensive set of experiments, totaling over 15,000 trials, wherein we meticulously compared the performance of various GNN training methods, focusing on popular models like GraphSAGE, GAT, and GCN. Our investigation delved into crucial factors such as accuracy and training time, with a specific emphasis on Multi-GPU distributed GNN training, employing cutting-edge frameworks like PipeGCN, DGL, and Quiver (PyG). To ensure a well-rounded analysis, we also harnessed diverse sampling techniques, including Neighborhood, Cluster, and SAINT, implemented using PyTorch on AWS EC2 instances equipped with GPUs. This research significantly contributes to the evolving landscape of GNN training methodologies, shedding light on the trade-offs between full graph and mini-batch training strategies while advancing our understanding of efficient and scalable GNN training processes.
Feature extraction from 3D CAD models
Under the guidance of Prof. S.S. Pandey at IIT Bombay, I conducted a remarkable research project focused on feature extraction from CAD models using Convolutional Neural Networks (CNNs). The project involved a multi-step approach to identify and represent 3D freeform features in STL format. To begin, I applied the Density-Based Clustering (DBSCAN) algorithm, a powerful tool for segregating these intricate features. Next, I harnessed the Speeded-Up Robust Feature (SURF) extraction technique to obtain feature representations in the x, y, and z directions, creating a comprehensive feature tensor. Leveraging this rich data, I trained a Multi-View 2D Convolutional Neural Network (CNN) using labeled feature vectors as inputs. The culmination of this effort was impressive, with the model achieving over 95% accuracy in automatically identifying freeform features within CAD models. This research showcases the potential of advanced techniques in enhancing CAD model analysis and feature extraction.