Edge Executive Interview – Myungsu Chae, Nota

MC: Nota was founded in 2015 based on the idea of ​​resolving minor inconveniences in everyday life through AI. Four co-founders who majored in deep learning started this business to create a practical tool that can affect people’s daily lives, leaving the research center where we mainly conducted advanced research.

The start of the company was “Nota Keyboard, a DL-based anti-typo keyboard that reduces typos.” This application provided customized keyboards to individuals by analyzing the user’s keyboard touch points and conversation patterns to learn typo patterns according to hand shapes and habits. After developing Nota keyboard, we aimed to operate DL on smartphones to solve privacy infringement issues, and ultimately, it has evolved into an on-device form. 

Starting from this, we developed DL model compression, the key technology optimized for Edge AI. Currently, we are expanding our core technology developed in-house to be deployed in diverse industries including Mobility, Security & Surveillance.

We are confident that the DL model compression of Nota will greatly contribute to overcoming limitations such as personal information issues, server and cloud construction costs, and communication restrictions that must be supplemented for various industries, everyday use of AI.

What is it you & your company are uniquely bringing to the edge market?

MC: Nota developed NetsPresso which is a unique solution for HW-SW co-design and it is the complete approach for edge AI. NetsPresso is a platform that automatically lightens and transforms vision-based DL models to meet the requirements of target edge devices (accuracy, latency, memory, power consumption).

Although there may be other similar platforms, Nota’s NetsPresso is the only truly device-agnostic platform that requires no specific chipsets or devices, and still covers all the available compression methods (i.e., pruning, quantization, filter decomposition, knowledge distillation, etc.).

At the same time, it is possible to create a model that can be operated on existing CPUs and memories without HW upgrade or design change. This also covers areas that existing model development competitors cannot support.

Additionally, NetsPresso’s Device Farm has the function that recommends a device suitable for the Inference time desired by the customer. NetsPresso’s Model Search has features that help customers with labeled data automatically create and deploy models themselves.

Unlike other companies that are only proficient in R&D or commercialization and lack the latest technology, Nota has established its own internally developed core technology and infrastructure, making it quick and easy to commercialize the R&D-level model compression technology. Nota is ready to respond quickly to changes in the marketplace since we offer various customer needs more efficiently than others by using our core technology.

Based on NetsPresso, Nota has established strong partnerships with Nvidia, Intel, and ARM to realize the ultimate HW-SW co-design.

Tell us more about the company, what’s your advantages compared to others on the market, who is involved, and what are your major milestones so far?

MC: Nota commercialized various vertical solutions (lightweight DL solutions) based on the horizontal approach that extends the original technology. The big difference from other competitors who simply lightens DL models is that we approach the market with a different business model vision.

Compared to HW accelerator, another alternative for Edge AI, we are sure that the DL model compression technology of Nota will be more prevalent in HW during the current semiconductor crisis, allowing more functionality and faster inference.

Platforms handling DL require customization based on specific data, devices, tasks, etc. Nota created a vertical solution based on the horizontal platform (NetsPresso). Horizontal platform and vertical solutions reflect feedback between each other and enable continuous reinforcement. We call this “the virtuous cycle of Nota” and it’s one of the strengths.

In addition to Nota’s Korean headquarters, we are expanding the business in Berlin, Germany, and North America, CA with our local offices. Currently, there are 60 passionate and competent members in Nota globally, and more than 70% of the company’s personnel are machine learning majors, which shows that our company is a technology-intensive company working in R&D.

Based on the infrastructure we have got, Nota successfully commercialized the NetsPresso compression in July 2021. NetsPresso’s Model Search and Device Farm will also be released at the end of 2021, and in the first half of next year, we will release full-version NetsPresso with full pipelines.

Nota’s NetsPresso is gaining attention in the global market as it begins commercialization of global IT and manufacturing companies. Nota’s vertical solutions, lightweight facial recognition solution (Driver Monitoring Solution, DMS) and lightweight intelligent traffic control solution (Smart ITS), are scheduled to be commercialized later this year.

DMS will push for mobility OEMs based on PoC success with Hyundai Motor Group in the second half of 2021. Smart ITS will expand its scope by participating in several large-scale smart city projects (ITS California, ITS Dubai, etc.).

How do you see the edge market developing over the next few years?

MC: Nota started edge AI very early (in 2015). We have personally experienced growth since the beginning of the edge market until now. In the early days, attention was focused on cloud and server-based AI, but from 2018-19, many IT frontier companies have recognized the importance of edge AI, which is now explosively emerging. As several companies, including Nota have successfully established a reference by attracting investment in edge AI, the frequency of experiencing edge AI in industry and daily life is increasing.

In particular, with the rapid growth of AI semiconductors, co-design of HW and SW is becoming more critical. Nota provides the platform that can perform the best efficient SW-stack, HW compatibility, and support for various functions. At the same time, we are strengthening partnerships with DL SW companies, including Nvidia, Intel, ARM, AMD, etc., and sometimes even M&A. 

As Smart devices (i.e., AI box) increase, we believe that increased global privacy, initiated by the EU GDPR, will accelerate the use and generalization of edge AI in policy terms.

What are the main trends you see about integrating edge computing into different verticals?

MC: The core of the main trend is HW-SW co-design. Until now, HW and SW were developed separately, but in order to maximize HW versatility and apply SW in various verticals, it is necessary to sufficiently experiment and test SW applications from the HW design stage. We believe it will be possible to control customers’ needs for the significant increase in semiconductor supply demand.

Recently, device companies (Nvidia, Intel, ARM, etc.) have been designing chips or devices in partnership with competent SW startups to take account of the shape in which SW-stacks work well. This isn’t easy to succeed unless it is based on a solid mutual partnership.

Therefore, through sufficient trust and technical verification, cooperation based on very strong partnerships will be an essential action strategy in the main trend. Nota is promoting similar types of cooperation with Nvidia, Intel, ARM, etc., based on a powerful partnership and is gradually creating a successful case.

As Elon Musk mentioned in his Tesla 21 Q1 earning call earlier this year, the most critical technology for full-self driving (FDS) will be DL compression since it is an important technology to handle all the complex operations in the vehicle. Nota has built success stories in several industries (mobility, coverage, retail, etc.) based on DL compression and we aim to expand our technology to enable even very complex AI systems such as full-self driving and AI assistants to run on the edge.

Thanks Myungsu Looking forward to hearing more from Nota at the event !