About this website
The purpose of this website is to navigate the projects and miscellaneous info about stuff spanning mostly my areas of interest that I've completed and wrote.
The purpose of this website is to navigate the projects and miscellaneous info about stuff spanning mostly my areas of interest that I've completed and wrote.
My name is Tan Sarp Gerzile and I am currently a computer science student studying at UC Berkeley. My academic interests span machine learning (unsupervised learning specifically), probabilistic systems, robotics, and vision. I've completed two internships at Capital Knowledge Resources at Ataol (London, Remote) and Unlu and Company (Istanbul, Turkey) focusing on data science and quantitative analysis. I am currently an undergraduate researcher at Berkeley Laboratory for Information and System Sciences, conducting research on reinforcement learning in multi-agent systems. I also try to remain academically literate on economic subjects. Apart from career and academia, I am a concert-level pianist and spend most of my free time studying it. During my remaining free time, I try to cycle and hike as much as possible.
This is a project to find the ideal offset between images taken with different filters to minimize some metric distance loss between the images, used mostly to colorize them.
Colorization ProjectThis project plays with different filters to examine the effects on a set of images. Image sharpening and blurring is used to create hybrid images. Finally, with the help of Gaussian filters, Laplacian Pyramid is constructed with examples.
Filtering ProjectThis project utilizes manual correspondace selection between two images of faces (in this case a child and an adult portrait of mine) and a Delanuay triangultion of these correspondances to generate two sets of Delenuay triangulations to calculate the midway face and find the affine transformation to it from the source image to warp the source image into the destination image using inverse warping. This method is then used to generate structural mean faces of population with examples through the Danish population data set. Finally, the average facial structure recovered from these triangulations is used to warp my face to the danish population. Finally, low rank approximation of faces is provided through the use of eigenfaces with Principal Components Analysis.
Warping ProjectThe first part of the project tries to recover homographies between images through manual correspondences and with these homographies, warped images and mosaics are generated with examples. The second part utilizes the harris corner detection algorithm to detect corners. Then, Adaptive Non-Maximal Suppression is used to uniformize the detectede corner points. Then RANSAC is used to approximate homographies. These homographies are then used to create mosaics.
Auto-alignment ProjectThe objective of this project concerns image generation with diffusion models. In the first part of the project, I use DeepFloyd UNet to generate text-based images based on the denoising process in the denoising of diffusion probabilistic models (DDPM) paper. In the second part of the project, I implement the necessary time and class conditioned UNet, training with the MNIST databased. This model is capable of denosing very noised hand drawn digits with the same process as the DDPM paper.
Diffusion ProjectBlog containing articles regarding math, economics, personal opinions, etc.
Personal BlogThis research explores the implementation of policy iteration algorithm in Q-learning to develope an optimal central policy for the multi-agent system.
Research LinkImplementation of the Pintos operating system by implementing userprograms, multithreaded programs, and a feasible FFS file system for the operating system.
Pintos LinkImplemented a fault tolerant MapReduce system in C by creating a coordinator process that distributes tasks to worker processes that carry out the actual computation. This system also handles worker failure by task redistribution and the coordinator communicates with workers using remote procedure calls.
MapReduce LinkDeveloped a dynamic stochastic general equilibrium model for asset pricing using GMM. Designed the model to incorporate economic indicators, performing parameter estimation with GMM, and validating against historical data.
GMM Link