Siddhant Garg

I am a Computer Science Master's student at the University of Massachusetts Amherst (UMass Amherst). My studies are focused towards Machine Learning and AI. Here I have worked on 2D/3D Computer Vision, Natural Language Processing and Optimizing Deep Neural Networks for GFLOPs (computation) reduction.

I completed my undergraduate studies from the Department of Mathematics and Scientific Computing at the Indian Institute of Technology Kanpur (IIT Kanpur) where I worked on Bayesian Statistics.

After that I worked with Samsung Research, Bengaluru, India for 2 years as a Machine Learning Engineer for On-Device AI solutions, where I published a paper as well as developed a deep learning models for novel applications to be deployed on smartphones globally.

During the summer of 2022, I also worked as Research Scientist Intern at Adobe Research and worked with the Video Understanding Group.

Resume  /  LinkedIn  /  Github  /  Google Scholar  /  Email

We present a deep learning method that propagates point-wise feature representations across shapes within a collection for the purpose of 3D shape segmentation. We propose a cross-shape attention mechanism to enable interactions between a shape’s point-wise features and those of other shapes. The mechanism assesses both the degree of interaction between points and also mediates feature propagation across shapes, improving the accuracy and consistency of the resulting point-wise feature representations for shape segmentation.

Deep Multi-Task models can be further optimized via model compression. In this work, we investigate the effectiveness of structured pruning on multi-task models. We show that, with careful hyper-parameter tuning, architectures obtained from different pruning methods do not have significant differences in their performances across tasks when the number of parameters is similar.

This work was done at Samsung Research, Bengaluru. We proposed a low-latency Self-Attention MobileNet for real-time image tilt correction for smartphones. The model was also deployed on the Samsung Galaxy with One UI for the Tilt Correction feature on Samsung Smart Share Tray.

We proposed a self-supervised method that learns the embeddings by reconstructing the 3D point clouds from noisy data with a Transformer encoder-decoder architecture.

We built a distributed video processing systems using machine learning for large video corpus with PySpark, CLIP and faster inference, distributed storage and faster retrieval.

We proposed a self-labeling refinement method for self-supervised model called Bootstrap Your Own Latent (BYOL) for more robust representation learning.

We used multi-lingual hate speech dataset called OLID and applied various transformer models to identify the best approach for detecting hate speech across many languages. We have used a multi-lingual transformer model called XLM-RoBeRTa, and we have also used BeRT models pretrained specfically on individual languages.

Implementation of Proximal Policy Optimization (PPO) for Lunar Lander and Cartpole.

Proposed an Expectation Maximization algorithm to find the parameters of the GSN distribution with the LASSO and Elastic-Net regression objective. The learned distribution is multi-model and the regularization terms help in modeling high-dimensional datasets with very few data points effectively. For example, we were able to find the most representative covariates, responsible for cancer prediction in the ARCENE dataset.

Devloped a proposal distribution for using Metropolis-Hastings algorithm to estimates the parameters of the Geometric Skew Normal Distribution.

Design and source code from Jon Barron's website.