Bayesian Methods For Machine Learning
This post will be a summary of my understanding from the coursera course Bayesian methods for Machine learning.
About Me
Currently, I’m working as a 3rd year PhD student in Florian Jug’s lab. My PhD project is an image decomposition task of splitting a superimposed image into its constituent channels.
I received B.Tech+M.Tech in Computer Science in 2015, from the IIT Delhi. I worked as a Data Scientist at Qplum, an online investment advisory firm, from Dec 2015 to Dec 2018. I then worked as a Research Assistant in National Taiwan University, Taipei under Prof. Hsuan-Tien Lin where I worked in Computer Vision.
PhD
Given a superimposed (noisy) image, the task is to decompose it into its constituent channels.
We published uSplit, a HVAE inspired architecture which used surrounding context in a GPU-efficient way, at ICCV 23. Subsequently, we built on top of it to create denoiSplit, a network which could do unsupervised denoising together with decomposition. This enabled (a) getting multiple predictions (a.k.a opinions of the network) for a single input and (b) model calibration: having a pixelwise estimate of error in the prediction. We recently published it at ECCV 24. For a meaningful evaluation of our work, and in general for unsupervised denoising tasks on microscopy data, we developed microSSIM, a variant of SSIM more suited for microscopy data which was published at BIC workshop, ECCV 24. At the beginning of my PhD, I also worked on structural noise removal using contrastive learning on the latent space.
Code Contributions
On the coding side, I’ve been interested in contributing to microscopy data related code: (a) c-mda-engine, a proof-of-concept C++ based engine for controlling microscopes. Idea was to move re-usable code from different downstream applications to a common codebase and use SWIG to generate python bindings. (b) I recently (May 2024) started contributing to microsim, a light microscopy simulator. (c) I created a python package predtiler, a light-weight package which enables tiled prediction in any dataset class used with pytorch with minimal changes in the dataset class code.
RA in Vision-lab, NTU, Taiwan.
Prior to this, I was working as a Research Assistant in National Taiwan University, Taipei under Prof. Hsuan-Tien Lin. There, I published two first-author works:
- 3D Gaze estimation in the wild. BMVC 2021.
- Extreme precipitation prediction. Problem Statement, Paper Link.