Working on the Automated Sign Language Detection System

Worked on the Unsupervised Compound Domain Adaptation for Face Anti-Spoofing under Dr. Luc Van Gool. Paper under review in IEEE Internation Conference on Automatic Face and Gesture Recognition 2021

Teaching Assistant for "Machine Learning for Natural Language Processing 1" in Fall Semester 2020. Responsible for taking tutorial lectures, clarifying Q&A on the online forum, prepare and evaluate exercises from basic (Logistic Regression for Text Classification) to advance level (BERT using Huggingface library), share useful study materials and provide technical assistance.

To Develop complex Machine Learning algorithms for ShoeBOX audiogram classification. Preprocessed Project Data (Audiograms) to identify data quality issues and outliers. Also, created Dynamic Computer Vision Algorithms to correctly identify different “markers” in over 4500 Audiograms. The Algorithm successfully detected the locations of the “markers” using OpenCV for the further extraction of information from Audiograms. Accuracy for parsing the data from Audiograms was nearly 85%. Above information from parsing the data from Audiograms and other provided features of Patient Records by the Children's Hospital of Eastern Ontario(CHEO), Ottawa was used to train predictors for the encountered diagnoses in the ears.

Implementation of Convolutional Neural Networks on FPGA. Convolutional Neural Networks have shown a lot of potential in real-time computing because of their low power consumption and high accuracy. One such application is in the field of image recognition and detection. This project focuses on the use of ConvNets for Numerical Detection, its software implementation using Python using Caffe, and its Hardware implementation on the FPGA. Current solutions, however, are based on using a group of GPUs for the computations. This leads to high power consumption. In order to reduce the power consumption while achieving realtime computations, researchers have started to do research on implementing neural networks on FPGAs. The traditional LeNet architecture of ConvNets, using the ReLu activation function was implemented, using the MNIST Database of more than 60,000 grayscale 28x28 images to train the network. This way, the accuracy achieved was nearly 99.98 percent.