“When a new function is required, who will research it within the company, and how will it be verified?”
- Interview with Hancheol Park (PM) and Shinkook Choi @performance team
NetsPresso is a proprietary platform that optimizes AI models on devices equipped with artificial intelligence technology. As an all-in-one platform capable of digesting multiple technologies, Nota has assigned several teams to develop and support NetsPresso's ongoing efforts. Nota's performance team is responsible for advancing and expanding the range of functions that NetsPresso provides.
As can be surmised from its name, the performance team is responsible for reflecting practical feedback in relation to the functional performance elements of NetsPresso. Exactly what does this work consist of, and who is the team behind it? We learned more about this through an interview with Hancheol Park, the team's PM, and Shinkook Choi.
How would you describe Nota's performance team in one sentence?
The performance team works to support NetsPresso, making it stronger, and ironing out any kinks that may show up from time to time.
Tell us more about the performance team. How is it structured?
[HC] The performance team comprises a total of seven Deep Compression Research Engineers, and is divided into two closely related camps: the research track, and the engineering track.
Two subteams? That's an interesting approach. What does each track do?
[SK] The research track is responsible for advanced research, focusing on lightweight deep learning models and efficient model designs across various fields, including, but not limited to, image classification and object detection. Our aim is to improve practical application, it is not merely academic research. We are primarily focused on developing new technologies and techniques. Once a new lightweight pipeline has been created and coded, it moves into the practical stage after verification with the engineering track.
[HC] The engineering track verification ensures that the lightweight pipelines and models created by the research track are suitable for use in NetsPresso. Additionally, as a team of skilled engineers familiar with the platform, we provide specialized lightweight modeling services to our customers, and continually work to improve NetsPresso in line with our customer's needs and desires.
You are currently on the lookout for a new Deep Compression Research Engineer. What does this role entail?
[HC] The successful candidate will have the opportunity to develop deep learning models using the latest compression techniques, in cooperation with the corresponding data sets for AI services. In short, they will have the opportunity to help us improve the totality of our on-device AI implementation process.
[SK] There is always room for improvement, and we are anticipating the feedback of NetsPresso users. Improved performance and supplementation is always high on our list of priorities. We want to satisfy the needs of our customers, and the successful candidate will work to make this a reality. It is an exciting position with a wide range of responsibilities.
[HC] To add a little more, there will always be room for personal development, and we are always open to new creative ideas. If an employee develops something new, or suggests upgrading existing technology, and it can be demonstrated to work with our products, we are happy to incorporate it into our arsenal.
I'm curious, what does a typical project look like for the performance team?
Project 1. NetsPresso Super Resolution Best Practice.
[SK] Not so long ago, we worked on a project with NVIDIA Jetson series devices. We had to lighten SR (Super Resolution1)) models in NetsPresso, something which had never been done before. The project went well and was even chosen as a NetsPresso Best practice. The NVIDIA Jetson family of devices are used throughout the world for a variety of AI models, so to have such success with this widely respected and popular technology is something we are very proud of.
While working on this project, we discovered a new, previously unknown, technique to lighten the weight of the SR model. Through this discovery, we were able to improve the NetsPresso Model Compressor2), which resulted in a new SR model with improved latency compared to the original baseline AI model.
1)Super Resolution: A technology that is able to create high-resolution images from low-resolution images.
2)NetsPresso Model Compressor: A NetsPresso module that efficiently lightens AI models.
Project 2. OCR (Optical Character Recognition) project with the company N.
[HC] In 2021, the performance team conducted an OCR project with N, a global gaming company. OCR technology is able to read text in pictures. For example, if you take a photo of your ID in a scenario where authentication is required, OCR is able to automatically enter your details, filling in the blanks of the form.
The project created a lightweight detection model for character recognition, optimized for NVIDIA Jetson devices. It was a company first, where we responded to an external customer using the NetsPresso Model Searcher1).
Thanks to NetsPresso, we were able to reduce latency by more than 20% compared to the existing baseline AI model relatively easily. At the same time, we were able to create a model that improved performance by 0.5%, to a model that improved latency by 63%, with minimal performance degradation.
1) NetsPresso Model Searcher: One of NetsPresso's modules that automatically finds an optimized AI model for a target device.
Both projects sound interesting. How are you working to expand and improve projects in the future? What skills are most important to this end?
[SK] The performance team serves as an internal NetsPresso user, so the ability to quickly learn and adapt to changes is paramount.
[HC] Lightweight modeling is an evolving field, with no standardized answers as yet. The things we learn and the experiences we have through hard work and diligent experimentation, will one day provide the basis for a more objective understanding. Until then, perhaps the most important attribute is a fighting spirit and a desire to achieve great results.
Working with verification is highly sensitive. Responsibility and decision-making skills are important. How do you foster a working culture within your team that embodies these qualities?
[HC] We follow a 'one person, one project' principle, so I have a lot of faith in the independent decision-making of our team. Additionally, I try to set aside some time for people to share their thoughts and feelings about work. At the moment, we have the following initiatives;
Six-week project retrospective: Undertaken with the philosophy that 'failure can be an asset', staff meet to deliver technical reports to one another, giving and receiving feedback.
Team Meetings & Coffee Talk: We have regular team meetings to discuss work progress, and to share opinions. Coffee Talk is a more laid back affair, where the team gets together for a casual coffee, with no strict topic or expectation in mind.
Team Seminars: We hold regular seminars on a variety of topics. This is the time to provide insights and ideas to everyone, covering research related to ongoing projects, or emerging trends.
As Project Manager, you made a promise to your team members. What was it?
[HC] When I became project manager, I promised my team that I would create more opportunities to collaborate within the company than anyone else. Given that the performance team is responsible for important research and verification of Nota's core technology, there should be as much collaboration between our teams as possible.
It is easy for teams to become isolated, and no one wants to be the frog in the well. We try to work together with other teams wherever possible, collaborating on research topics that crossover departments. This applies to all levels, from team members to executives. By working together, every team naturally aligns and the company is able to move forward into the future with a united vision. This is a key part of our success.
Speaking as members of the performance team, how would you assess your ongoing efforts?
[SK] I think the charm of our team's work lies in our ability to adapt. We enjoy working across a wide range of fields, using popular widespread tools, to new and emerging technologies. Whatever the project, we have fun discovering the challenges that face us.
[HC] I think there is a great sense of achievement to be found when you can meet a practical demand, rather than simply doing research for the sake of research. The performance team prides itself on being a team that considers the individual user, and their wishes, the most. When your vision aligns with the direction of the company, it is a perfect fit. Finally, it is rare to have an opportunity to deal with a such a wide range of technologies, and to feel that you are all working together to the same end.
Lastly, what would you like to say to those who might be interested in working for the performance team?
[HC] The job titles of compression engineer, or research engineer, might sound a little unfamiliar in the AI field. But if you have the experience and courage to begin modeling with deep learning frameworks, the job is challenging and rewarding, and you will be surrounded by a great team!
[SK] Are you someone who likes to research cutting edge technology? Someone who also likes engineering? A personal who wants to solve practical problems? The performance team is waiting, we can solve the world's AI problems together.