Вы находитесь на странице: 1из 2

Statement of Purpose

Pengxiang Cheng It has long been my dream to understand what intelligence is and how to make intelligent machines. As a senior majoring in Automation at Tsinghua University, I have had several undergraduate research experiences that motivated me to pursue a doctoral degree in machine learning. The choice arose from a long period of intrinsic reection on my interest and capability, yet the chief motivation is that I do enjoy research, especially the sense of fulllment from turning an imaginary idea into implementation. At the beginning of sophomore year, I took part in Tsinghua University Electronic Design Contest, which concerned designing and building a small intelligent car to accomplish tasks of chasing and evading each other. I led my team enter the nal round, and ranked top 20 among over 100 teams. However, the poor performance of infrared and photoelectric sensors to recognize surrounding objects of the intelligent cars encouraged me to investigate more ecient recognition methods involving vision techniques. Therefore I joined a student technology team, and led a research group on computer vision topics. Our rst goal was to implement a human tracking system with overhead cameras. Considering the performance of hardware, I decided to use simple Hough Transform to detect human head followed by CamShift object tracking algorithm, and eventually accomplished the task perfectly under non-dense circumstances. This whole-year experience of independent exploration in intelligent systems greatly enhanced my technical skills and scientic ambitions. With a strong interest in computer vision and solid foundation in statistics and programming, I joined Prof. Qionghai Dais lab in my junior year, collaborating with a postdoc on two projects. The rst project aims at building highly ecient refocusing cameras. Previous research used programmable aperture to capture light eld, causing the time cost proportional to angle resolution. We proposed to accelerate light eld capturing by reducing the number of shots and reconstructing missing views via optical ow estimation. I implemented the algorithm after four months of repetitive testing and modifying on real datasets, which signicantly improved eciency. Later I moved to another project, still using programmable aperture, but focusing on image relighting. The idea was to use coincident patterns on both multiplexing viewpoints and directional illuminations to capture sequential images, and then applied sparse decomposition to separate illumination components using the method of augmented Lagrange multipliers. I designed and implemented the optimization algorithm all by

my own, now the project is still undergoing. My eorts throughout the year greatly deepened my understanding of machine learning, also made me realize the learning process itself, especially the optimization technique, is what truly appeals to me. In the summer of 2012, I obtained the opportunity of internship in the Virtual Reality Lab at The University of Texas at Austin, under the guidance of Prof. Dana Ballard. I conducted an independent research project that focused on mathematically modeling muscle movements in human gaits. I proposed to use multi-channel sparse decomposition to substitute state-of-the-art coding techniques by simultaneously learning dierent muscle systems. This novel learning method signicantly improved the coding eciency and satised the inner muscle synergies including left-right alternation and exor-extensor alternation. Now we are making the nal revision of the paper and preparing to submit it to ICML 2013. Nevertheless, what impressed me most at UT Austin was not just the research but what Prof. Ballards attitude about research, You should enjoy research, dont give yourself any pressure. His words reminded me of my high school days, when I spent more than 30 hours every week to advanced learning in mathematics and physics without assistance on topics such as number theory, combinatorics and quantum physics. I loved the latitude of learning and discovering intriguing areas without any reliance, and the satisfaction from solving problems independently. I was enjoying the latitude and satisfaction again in Austin, just as Prof. Ballard said, which conrmed my resolution in pursuing a PhD, a long period of independent research, and a way to professional in the area of machine learning. The most exciting part of machine learning to me is its possibility to help us understand human learning and the way brain works. Some relevant progresses have been made, for example, the sparse representation of images with visual cortex features, and more recently, the incredible Google Brain project. However, there is still much work to be done to gure out the computational approach under intelligence, which shall be the goal of my life. For now, I am interested in two specic topics. The rst one is using online learning to simulate human visual attention. A possible solution is to use eye tracking equipment to record visual attention in real circumstances, and then use this record together with visual information as auxiliary guidance for statistical learning, just like teaching a child to see the world. The second one is about deep learning, which has been drawing much attention in past few years, but only a few successful implementations exist. I am eager to explore its applications in traditional machine learning topics.

Вам также может понравиться