Profile Picture
Bhuyashi Deka
Computer Science Graduate Student

Hey,

I am a second-year Master's student in Computer Science at the University of Wisconsin-Madison. Currently, I am working in the RT2 lab under the guidance of Prof. Mike Hagenow, where my work focuses on Partial Observability in Imitation Learning for xARM robots. Previously, I worked in the WISION Lab under Prof. Mohit Gupta, simulating LiDAR sensor capabilities for SPAD cameras. My research interests broadly lie in Computer Vision, Robotics and Embedded Systems.

Prior to my graduate studies, I was part of the Multimedia and Algorithms team at Samsung Semiconductor India Research. There, I contributed to the image post-processing pipeline for Exynos SoCs, specifically developing the Dense Optical Flow and Spatio-Temporal Noise Reduction engines.

Before that, I completed my Dual Degree (B.Tech and M.Tech) in Electrical Engineering from IIT Bombay. For my Masters thesis, I was supervised by Prof. Virendra Singh where I worked on building efficient hardware accelerator for convolutional neural networks.

In my free time, I enjoy reading and painting. Please feel free to email me with questions or feedback regarding any of my projects.

Score 0
Episodes 1
Learning Rate 0.1

Projects

Variational Autoencoder Image

Variational Autoencoder for MNIST and Fashion-MNIST

Developed a Variational Autoencoder model for MNIST and Fashion-MNIST datasets. The model explores the latent space of the data and reconstructs the images from the latent space.
Domain Adaptation Image

Unsupervied Domain Adaptation for Semantic Segmentation

A test-time domain adaptation technique for semantic image segmentation that improves model performance on out-of-distribution data (like foggy cityscape images) by dynamically adapting batch normalization layers without additional training.
FCOS Image

FCOS: Fully Convolutional One-Stage Object Detection Model

Developed FCOS architecture that eliminates anchor boxes and provides efficient, accurate object detection. It explores object detection across multiple feature pyramid levels using ResNet backbones (ResNet18 and ResNet34), achieving a mean Average Precision (mAP) of up to 0.61 on test data.
Diffusion Image

Diffusion Model for Image Synthesis

Developed an advanced deep learning model implementing UNet architecture with diffusion techniques to generate synthetic images across MNIST and AFHQ datasets. Explored innovative techniques like sinusoidal time embeddings, cross-attention mechanisms, and latent diffusion models to reconstruct images through iterative denoising processes.
Donut Image

Creating a Realistic Donut in Blender

In this project, I created a realistic 3D-rendered donut in Blender, focusing on procedural modeling, texturing, and lighting. It showcases my work on detailed shading, particle-based sprinkles, and high-quality rendering using Cycles.
Global Clustering Image

Global Clusters: Hierarchical Analysis of Socioeconomic Indicators

The project uses Hierarchical Agglomerative Clustering to analyze global socioeconomic data, grouping countries based on metrics like GDP, literacy rates, and life expectancy. The results were visualized through dendrograms and scatter plots, offering insights into the clustering patterns.
Teeko Image

AI Teeko Game Player

Developed an intelligent AI game player for Teeko, a strategy game played on a 5x5 board. The project involved implementing the minimax algorithm with depth-limited search, heuristic evaluation, and custom scoring functions to simulate strategic gameplay.
Faces PCA Image

Face Reconstruction with PCA

Implemented Principal Component Analysis for facial image analysis to reduce dimensionality while retaining critical features. The project explored eigenfaces, variance preservation, and reconstruction to enhance understanding of feature extraction in high-dimensional datasets.
Pathfinder Image

Exploring Pathfinding Algorithms

The project provides an interactive visualization tool that demonstrates pathfinding algorithms like A*, DFS, BFS and visually compares algorithm efficiency.