Writing Feedback / Statement of Purpose on Computer Science, PhD [3]

Hi everyone, I am a master student who is applying PhD program in computer science.
I'm not a native speaker, and I am not quite sure whether the content I wrote is suitable for sop.
Could someone who has experience in sop writing help me judge two things? 1) Is this sop logically correct, like the things I wrote is exactly what is needed in a standard sop? 2) Are the writing clear and fluent enough? I'm afraid some writing is not natural and hard to understand by reviewers.
Thanks for everyone's help! Here's my sop:

My decision to pursue graduate studies stems naturally from my experience as a master's student at xxx University. During my time there, I developed a strong interest in generative models and controllable visual generation, with a primary focus on image/video generation, manipulation, and their interaction with other modalities such as text and 3D. I am particularly interested in the problem of learning efficient 3D representation from large video models. Through graduate studies at yyy University, I wish to delve deeper into this topic, but I am open to trying new ideas as well.

My academic journey began with studying Generative Adversarial Networks (GANs) and their applications. When diffusion models emerged, I noticed their Markov inference process, though effective for generation, could lead to error accumulation when used for image editing. Based on this, I proposed xxx [1], which integrates a well-designed hyper-network and an optimized training algorithm to reduce error propagation in editing process. This work was then published in [Conference XXX]. As my first research endeavor, it taught me great lessons of how to conduct research. One key lesson was not to take action only after everything is perfect. I used to overthink whether my ideas would work rather than conducting quick experiments to test them, and have dismissed many promising ideas for fear that they were not perfect. Since then, I have developed a fast research pace through trial, error, reassessment and iteration, which I consider to be crucial skills for conducting research.

During my master's studies, I've also engaged in projects across various domains. I led an AI4Science project on protein sequence generation. This is a biological problem that can be modeled as an NLP task, a field in which our lab had nearly no experience before. Thus, I self-studied NLP and independently solved the issues I encountered in papers and codes. After considering the unique properties of proteins, I developed a generative framework based on their hierarchical structure and unidirectional linking patterns. The proposed framework successfully generated proteins not found in nature, but with enhanced functionality than natural ones (This work is ongoing). In addition, I have studied reinforcement learning and applied it to optimize sequential decision-making tasks in practical scenarios [2]. I enjoy learning knowledge from other domains, for they provide new perspectives on my current work. Well-established areas like NLP, in particular, can offer valuable insights transferable to other fields. These projects have helped me develop the ability to learn quickly about new areas, as well as transfer existing knowledge to different contexts.

Most recently, I undertook a six-month internship at zzz company, where I worked on the controllable generation of large video models. In collaboration with Dr. aaa, we discovered current large video models are incapable of performing general camera control - text can only affect content and style, while finetuning is constrained by the lack of sufficient and diverse camera-annotated datasets, limiting their practical applications, e.g., film generation with cinematography. To address these issues, we proposed xxx [3], a training-free approach that provides camera control for almost any video diffusion models. It outperforms finetuned methods in both generation quality and camera motion accuracy, and more importantly, achieves truly generalized camera control across diverse video contexts. We built a website to exhibit these excellent results. Through deeper exploration, we found this method has greater significance than merely controlling video's camera motion: Our experiments revealed that, by applying this approach, video models are able to generate dynamic, 3D videos ― through a completely unsupervised manner.

This finding is exciting, because it serves as evidence that there may exist a more efficient 3D representation inside video models for better 3D reconstruction, generation and especially, dynamic manipulation. From this perspective, video models can also be regarded as the largest resource of 3D data as it should be, since most of our real 3D world is recorded through videos, rather than other forms like point cloud. It is a core problem behind several long-standing challenges in computer vision, such as unified 3D representation, 4D data modeling and world model construction. It could also influence the way humans interact with future XR technologies, and we might no longer need physics engine to model virtual world. I believe this is a topic which will have profound impact on the field of computer vision. And during my PhD studies, I aspire to explore deeper into these or relevant problems, as well as develop solutions that contribute to 2D/3D/4D communities.

yyy University has a distinguished faculty renowned for their exemplary contributions in the domain of computer vision. I would especially like to work with Prof. Dr. aa, Prof. Dr. bb and Prof. Dr. cc. Over my master's studies, I have been influenced by several of their works on 3D generation and representation learning, an opportunity to work with them is extremely appealing to me. I believe my research experiences, and more importantly, my insatiable desire to learn and solve problems, have prepared me adequately for dealing with the challenges in PhD journey and making further contribution to my research fields. Studying at yyy University as a PhD would undoubtedly lay strong groundwork for my future goal as a researcher.