온디바이스 AI 최적화 솔루션으로 업계 주목을 받고 있는 주식회사 노타(이하 노타)의 조석영 매니저를 만나 관련 기술과 업계 동향, 전망 등에 대해 이야기를 나눠봤다. 이어서 노타의 AI 솔루션 역량과 온디바이스 AI 기술의 미래에 대해서 심도 깊은 답변을 들을 수 있었다.

▲Jo Seok-young, Nota Manager
Nota, ITS Success Story Introduction... Latest VLM Technology Incorporated
Existing AI, Edge Case Response↓·'Context Blindness' Without Learning
Generative AI and Multi-Modal, Evolving into AI That Knows 'Context'
We met with Manager Seok-Young Cho of Nota Co., Ltd. (hereinafter referred to as Nota), which is attracting attention in the industry for its on-device AI optimization solution, and talked about related technologies, industry trends, and prospects.
(Part 1) Next, we were able to hear in-depth answers about Nota's AI solution capabilities and the future of on-device AI technology.
■ Nota's rapid growth is attracting attention. What is Nota's current status? This year's target sales are at least 10 billion won. Nota is a startup with rapidly growing sales. It is recording growth of about 2-4 times every year, reaching 500 million in 2021, 2 billion in 2022, and close to 5 billion last year.
Nota is planning a technology-specific listing, aiming for an IPO in 2025. Mirae Asset Securities has been selected as the lead manager and is currently gearing up for exchange review.
■ I am curious about use cases that utilize Nota’s solutions. The greatest strength of Nota's optimization technology lies in its versatility. It has the technology to optimize various AI models, from existing CNN-based computer vision applications to recently popular generative AI.
Representative success stories include: △Intelligent Transportation System (ITS) solutions, △Facial Recognition solutions, and △Driver Monitoring solutions.
In the field of intelligent transportation system solutions, the company has the highest level of performance in smart intersection systems and emergency control systems, and is currently building and operating these solutions in major local government ITS projects such as Daejeon and Yeongcheon.
In addition, the driver monitoring solution was supplied to a leading global black box company, and the facial recognition solution was supplied to a leading domestic home IoT company, respectively, and their technological prowess was recognized.
A representative example in the field of generative AI is the collaboration with Samsung Electronics. At MWC2024, the world's largest mobile exhibition held in Barcelona in 2024, Nota, together with Samsung Electronics, presented the world's fastest on-device image generation stable diffusion demo running on the Galaxy S24.
■ Global chip makers are also supplying full-stack AI software tools. What are the differences between Nota’s solutions and these? First, the scope of support is broad. Most semiconductor companies provide development tools that are limited to their own hardware. However, in the actual market, various semiconductor processors are used, so when developing a solution, different software libraries and development tools must be used for each device.
This is a factor that further worsens the decline in technological competitiveness due to the already lack of AI experts. In addition, each semiconductor company focuses on specific applications, so the model support range is limited.
For example, ST and TI, which mainly produce low-end MCUs, are focusing on machine learning applications at the time-series analysis level rather than high-performance computer vision or transformer-based models. In contrast, Nota’s technology enables various AI applications to be easily optimized and installed on a wide range of heterogeneous hardware platforms.
Second, lightweight and optimized technology. Since its establishment in 2015, Nota has continuously researched and developed lightweight technology for on-device AI. Based on the accumulated original technology and abundant domestic and international patents, it provides differentiated performance from open source and third-party technologies.
A typical example is pruning, one of the main techniques for reducing the weight of AI models. The unstructured pruning or Sparsity technique, which is mainly used in open source or third-party development tools, replaces the weights of the AI model with '0' to reduce the model size, but the actual reduction in computational amount is limited, so the speed improvement effect is minimal.
In contrast, Nota's structural pruning technique achieves substantial speed improvements by directly removing channels from the AI model. This allows for the weight reduction of the model while minimizing accuracy loss.
.jpg)
▲Jo Seok-young, Nota Manager, predicted that the development and convergence of generative AI will lead to the birth of on-device AI that overcomes the limitations of existing AI.
■ What are the limitations of existing on-device AI? For example, in factories, defect detection using AI cameras is often done, and at this time, the camera specifications, setting environment, etc. must be considered in model learning. If the camera specifications change or the angle changes, the existing model will not work. Therefore, the demand of factories does not prefer individual AI solution companies, but rather seeks companies that provide all-in-one services from hardware to software.
In addition, it lacks the ability to deal with new defects. In order to detect new defects, existing models must spend money and time collecting and learning data again. In particular, it is impossible to deal with edge cases, but the development of generative AI in the future will provide the ability to deal with defect cases that have not been seen before.
■ How will the development of generative AI technology affect on-device AI? Artificial intelligence technology has recently reached a major turning point. The evolution from existing discriminative AI to generative AI has become an opportunity to dramatically expand the scope and possibilities of AI utilization.
In the past, AI was limited to relatively simple judgment tasks such as classifying given data and recognizing patterns. Now, AI can understand ‘context’ in a human-like way and comprehensively perceive situations to create entirely new content.
It is important to note that generative AI demonstrates flexible problem-solving capabilities. While traditional AI clearly has limitations in responding to new situations that it has not learned, generative AI has acquired the ability to solve problems through inference even in new situations where there is no existing learning data.
■ What are some examples of current generative AI technologies being applied on-device? In the case of Nota, this technology was applied to the intelligent transportation system (ITS) and achieved groundbreaking results. A solution was developed to automatically detect traffic accident scenes on the road that were previously difficult to detect in real time using edge devices. This technology attracted industry attention when it was demonstrated at NVIDIA's annual developer event, GTC, in 2024.
It is not just a vision, but also a language model, and it is a multi-modal vision language model (VLM), and Nota is internalizing and integrating this model into its solution. Through this, you can create accident reports and ask and answer questions if you have any questions.
Not only Nota, but also major global IT companies such as Google, Samsung, Apple, and Qualcomm are accelerating the development of relatively compact large-scale models such as small language models (sLLMs) to drive generative AI in on-device environments. This is an effort to bring the benefits of generative AI to general consumer devices rather than high-performance server-grade hardware.
■ A final word to the readers Technological advances will allow AI to penetrate deeper into our daily lives and industrial settings, taking user experiences to the next level, such as generating high-quality images in real time on smartphones or having more natural and contextual conversations with personal assistant AI.
In industrial settings, changes such as more flexible responses to unpredictable situations, creative problem solving, and taking automation to the next level are expected.
We ask for your support for Nota, which will lead the on-device AI era, and we also hope for your interest in AI optimization and lightweight technology that connects AI and semiconductors. We also welcome inquiries about collaboration at any time.
thank you