Research

2020. 11

I have a new website! Please visit here: https://joshualin24.github.io/

2020. 8

Extremely happy to share our new paper ” Feature Extraction on Synthetic Black Hole Images” that was accepted by the ML Interpretability for Scientific Discovery Workshop at ICML 2020.  link to the paper: arxiv.org/abs/2007.00794 

Here’s a short video we made to present our work: https://www.youtube.com/watch?v=D4QNmjXtp3Q

2019. 1

My research focuses on astrophysics and cosmology, especially on dark matter. Dark matter is a hypothetical massive matter that neither emit or absorb light (photons), so they are invisible to us. The only few ways to detect them so far has been through their gravitational effect on light rays from luminous matter (distant galaxies) according to general relativity, which is called gravitational lensing. By studying these gravitational lensing systems, we can understand the distribution of dark matter in our universe. I’ve been trying to use computer vision with deep neural networks to study the distribution of dark matter substructures in these gravitational lensing images, mainly via simulated gravitational lenses. Python is my strongest programming language. I’ve used python to build up the gravitational lensing simulator, mainly using numpy and scipy. I further used deep neural networks (ResNet, DenseNet) to do computer vision (supervised learning) on these gravitational lensing images. We are at the stage of publishing a research paper and we are all excited about this result.

Another exciting topic I recently started is doing Feynman’s Path Integral with reinforcement learning. I noticed that the principle of least action notion in classical physics is extremely similar to problems in maze-solving/optimal path planning in reinforcement learning, which we can treat “action” in the principle of least action as a reward/penalty. Feynman has extended this idea into quantum mechanics and quantum field theory. Therefore, we could build up a system that learns intuitive physics with reinforcement learning. My Goal for this project is taking intuitive physics into quantum level, and use it to solve tons of applied and fundamental problems in physics.