Namshik Kim

I am a data physicist, interested in deep learning. The posts of the blog demonstrate my interest very well. I still think of myself a physicist because my perspective in data science and deep learning is still very physicist-oriented after I finished PhD program in physics in string theory group in May, 2017.


Current Résumé

Data analysis

I do not talk about SQL or other data analysis topics. If you want to know about my skills on data analysis, check THIS.

Current interests and projects

I had worked for 1 year at Soundcorset which developed the application Tuner & Metronome. Two projects I participated in are

  • The convolutional neural network based machine classification project evaluates the clients’ performances and motivate them by presenting daily top performances. The genre/instrument classifier basically uses the same neural network model and it classifies the clients’ recordings and improved conveniences. This project contributed to app-download increase. The company is ambitious to extend the Soundcorset tuner & metronome into an ultimate music box for practice. This work is the first step for the goal and is applied to HERE.

  • I devoted to the a musical data generation project. It transfers clients’ recording to other genre of musics. I contributed to the proper data preprocessing (music to image ) and post processing ( image to music) and the neural network model construction based on recentest deep learning research. (GAN/MUNIT). I have explored possibility of spectrogram-image process using various GAN (generative adversarial networks) for music. The post of this blog presented Introduction to GAN.

Published papers

5 published string theory research papers can be found here.

What did I do in academy : Scientific model builder in string theory

During PhD program of physics, I have been building a gravitational model to understand strongly coupled quantum field theory by using duality suggested by string theory. Roughly, it was a trial to understand special phenomena can occur in double layered graphene. Graphene has amazing possibility in electronic applications, but it is extremely hard to figure out its electronic behaviors.

Why my work was not easy and what kind of skills I obtained : a serious analytic problem solver and also a numerical technician, and have statistical research background

To solve complicated equations of motions, scientists use various approximations. There are only few things to solve clearly in nature. If the interaction between particles is too strong, most of approximations are not valid. When you pay for a 124 dollars product at mart, to prepare a 100 dollar paper, a 20 dollar and 4 one dollars is the best way. Try to optimize from the biggest piece and then 2nd biggest and so on. This kind of differentiation is not at all working in the world of strong interaction.

For another example, modeling in physics can be compared with building house in some sense. If you have a foundation of the building, you would put some ingredients on it to complete up the building. In my research, the object of our interest does not have suitable tiny ingredients, so we need some miraculous duality can exchange the ingredient to the suitable.

By string theory, a gravitational theory can be replaced by solving above difficult problems. Solving a gravitational theory is still not so easy, but is well-known. The equation of the motion obtained from the gravity theory is highly non-linear and complex, so I had to analyze them numerically. Black holes are solutions of Einstein’s gravitational theory, and they are also thermal objects. Then, studying statistical mechanics is essential to research the models. I have numerically or analytically test my model and have found many statistical meaningful results including phase transition. It turned out all the skills I gained are very fruitful in machine learning.

Transition to machine learning

Therefore, my research in the string theory group was to investigate thermodynamical preference of the gravitational models which is considered to match condensed matter systems. The former sentence includes 3~4 different field in physics already. The biggest challenge in my PhD program was to understand many different topics as quickly as possible.

Physicists are familiar to solving the problems, and what data scientists do in machine learning is not at all different. The approach to the questions and how to solve them is very similar to what we did in string theory in some sense. Machine learning is already very broad subject and some of them are researched and elaborated by high energy physicists as well. Some of the graduate level machine learning textbooks are written by former high energy physicists, and when I read them, I feel very comfortable at their logical approaches to the subjects.

Above all, machine learning is fun.