Research Intern
University of Toronto
Total years of experience :3 years, 0 Months
Crafted synthetic material datasets to experiment with various generative models such as the Variational Autoencoders (VAE), Generative Adversarial Network (GAN), and the latent diffusion model.
● Trained a VAE to produce reconstructions of textures, resulting in generating never-before-seen materials while reducing data representation down to 2%.
● Single-handedly realized a novel loss function based on the Fast Fourier Transform (FFT) to produce reconstructions of textures that preserve the spatial statistics of the original texture while not necessarily reconstructing the same image in the data space.
● Using the FFT loss function, enhanced training speed by a factor of 2 and reduced GPU RAM requirement by a factor of 4.
● Tools used: Python, PyTorch, FFT, deep generative models, CNNs, VAE, transfer learning, latent diffusion models, generative adversarial networks, graph neural networks, attention mechanism.
Developed a semi-supervised machine learning algorithm leveraging Integer programming and deep learning (autoencoder) in Python that predicts the number of people in a house from WiFi channel state information (CSI) with 80% accuracy resulting in bringing an additional estimated $2M value to the company.
● Created an unsupervised machine learning model that improved motion localization and prediction in a house from WiFi CSI. The refined model employed clustering, integer programming, and Natural Language Processing principles (NLP-Bag-of-Words) and achieved 95% accuracy and resulted in bringing $2M value to the company.
● Applied mixture models, spectral model, and multidimensional scaling from scikit-learn to cluster noisy data resulting in an 80% accuracy score.
● Built a hidden markov model algorithm to extract hidden states (number of people in a house) from the noisy house occupancy data using integer programming resulting in an additional 15% improvement.
● Tools used: Python, Scikit Learn, Numpy, PyTorch, hmmlearn, Linux, Google OR tools; signal processing with autoencoders, and hidden markov models; Dimensionality reduction with autoencoders and PCA; manifold learning with Spectral models, multidimensional scaling, and mixture models.
● Enhanced a signal segmentation algorithm based on calculating rate of change in signal amplitude in Python which elevated segmentation accuracy from 40% to 85%.
● Applied kurtosis analysis algorithm to detect misclassified signals using statistical analysis; advanced performance from from 74% to 98%.
● Tools used: Python, Scikit Learn, Numpy, Pandas, Linux, signal processing using convolution.
● Contributed to a robust cost prediction model by leveraging a PyTorch-based neural network regression framework, tailored specifically for estimating production expenses at Calvin Klein during their commercialization phase. Achieved a prediction accuracy rate of 99%.
● Tools used: Python, Scikit Learn, Numpy, Pandas, PyTorch, SQL.
Master of Science in Computer Science Candidate