Statement of Purpose

Learning is an important process that helps every living being to evolve, develop, and resist being extinct. Among all, human shows the highest capacity to learn, in order of magnitudes greater than any other. However, one important issue that troubles me was “Could humans learn everything?” When cogitating about the massive universe with its enormous unexplored data, this is skeptical. In such a context, I believe that more work awaits for machine learning enthusiasts like me, in particular, to persuade related studies and contribute to the development of machine learning and artificial Intelligence.

The most impressive thing I found in machine learning is its’ wide spectrum of applications. As the application spectrum spans more and more, and algorithms become advance, the demand for computation power more than any other field. In addition, we can find machine learning operates in different environments; from supercomputers to mobile devices, from cloud computing to edge computing. These different environments possess different power, performance, and area requirements. This has fueled challenges as well as opportunities for researchers. In both academia and industry, we can see many efforts to solve these challenges: “Lightweight networks for edge devices”, “Semi-Supervised Learning to circumvent large dataset requirements”, “Reinforcement Learning to tackle unconventional problems” etc.

In the field of machine learning, we have user-friendly frameworks such as TensorFlow, PyTorch. Further, there is a significant interest in developing efficient hardware accelerators to run machine learning algorithms. But Hardware architects often face issues in adapting to applications written on TensorFlow or PyTorch in their hardware architectures. This gap between software and hardware layers has become a significant burden to many machine learning enthusiasts around the world to use machine learning efficiently in the real-world. To close this gap, efficient end to end compiler stacks are required. Therefore, the collaborative work of these three fields is required to win the machine learning revolution. I believe my experience which spans over these fields will be helpful in ongoing research in this regard.

I realized the importance of the above concept when I was completing my undergraduate research project. It was titled “Machine vision processor for leaf node”. The projects’ challenging nature and research orientedness encouraged me to choose the project. The target was to develop an application-specific processor using an FPGA, to run feature extraction algorithms. In the project, I mainly involved with RTL designing, developing software models, and do verifications. During the literature review phase, I found many fascinating works, which improve performance through efficient hardware architectures. However, the trade-off was programmability. In designing our hardware architecture, we faced the same issue. We knew that, to handle this problem we need to design an efficient compiler. We researched more on this direction for future improvements. We were able to successfully complete the project and published a paper titled “VLIW Based Runtime Reconfigurable Machine Vision Coprocessor Architecture for Edge Computing” at the 30th International Conference on Application-specific Systems, Architectures and Processors (ASAP) held in New York, in July 2019.

I knew that performance of the design could have been enhanced if a compiler tool-chain had been developed. This aroused my curiosity to question myself; “How the different Application Specific Processors are programmed?”, “How the mapping of high-level instructions to low-level is done?”, “Can we program it in a way that exploits the parallelism supported in dedicated hardware?” To find answers to these questions and to quench my thirst for machine learning, I joined Wave Computing.

This motivated me to choose my first job in Wave Computing Sri Lanka as a Software Engineer in the Software Development Kit (SDK) team. It is very rare to find a job in these fields in Sri Lanka. Wave computing was a silicon valley based startup. Even though, it was a research-based one, which is another reason to join the company. Being the youngest member of the team, I got the rare opportunity to work with amazing, well-experienced engineers in the industry. Our goal was to develop a software stack to program massively parallel CGRA architecture type processor. This processor was intended to use in machine learning applications. Working in the team gave me exposure to both machine learning and machine learning related compiler development.

In the journey, the next destination was Acceler, a startup focused on machine learning and the acceleration of software and hardware systems. However, at the start, my job description did not include machine learning or computer vision tasks. When I let them know about my interest in Machine Learning, they were happy to assign me to machine learning Projects. This gave me an opportunity to apply theories I have learned, to solve real-world problems. I worked with machine learning frameworks like TVM, Xilinx ml-suite, Distiller and I was fortunate to participate in a project on analog inference for neural networks. I used the opportunity to broaden my knowledge about cutting-edge-technologies that are currently being researched by researchers. During these projects, I exploited the wealth of resources related to machine learning and deep learning. In one of the project we were required to come up with a system for automatic draught reading and the contractor wanted to use machine learning to achieve the task. In this project, we built the dataset, designed a machine learning back-end system, and a front-end web app. I found this is experience valuable as this helped me to realize “Building the dataset” is the most important part of a machine learning project. This decides your models’ accuracy even before you come up with an architecture. “Data is indeed the new oil”. This sparked me to learn techniques to circumvent large dataset requirements. Particularly, semi-supervised learning is a research area that fascinated me. Recently I contacted one of my final year thesis supervisors to participate in ongoing research available at my alma mater, the University of Moratuwa and he had given me a topic related to Semi-Supervised Learning. Currently, I’m in the literature review phase and I’m confident that I can balance my work at the office and research at the university.

My application to Stanford University came after the discovery that the university has placed very significant research focus on machine learning. I am particularly fascinated by the work done at Stanford Vision and Learning Lab. I believe my experience in the industry and the academia aligns well with the vision and mission of the lab. I am also interested in the work of Professor Monica S. Lam on machine learning for virtual assistants. I am also intrigued by Professor Clark Barret's work on Satisfiability Modulo Theories (SMT). My experience with the "SAT solver" based compiler at Wave computing, aligns with his research.

My academic credentials which eventually ranked 2nd in the batch show that I have the required analytical thinking capability to engage in graduate studies. My academic performance coupled with my experience in both research and industry will help me to succeed in graduate studies. Work in industry enlightened me on how the theory is realized in the engineering world and what should researchers do to realize their work in the real world. During both academic and industrial projects, I self-learned the necessary knowledge, I have proven the possession of the enthusiasm and drive to engage in a research-oriented career. Having these qualifications, I confidently believe that, given the opportunity, I will be able to make a significant contribution to exciting research happening at Stanford.