DEVIEW 2019

Since 2008, DEVIEW has been the South Korea’s most prominent tech forum on software engineering. It is unarguably a forum from which developers and researchers share ideas and get inspirations. Approximatively 3,000 local and foreign software developers and tech industry officials participated in this year’s conference. DEVIEW 2019 was held at COEX Grand Ballroom, Seoul, from October 28 to 29.

President Moon Jae-In attended DEVIEW 2019 as a keynote speaker. His speech carried significance in that he was the first president to attend to the event and deliver the keynote address at such a prestigious software engineering conference. During his keynote address at DEVIEW 2019, President Moon Jae-in promised to provide significant support on the nation’s artificial intelligence. He said that the government had allocated an increased budget of 1.7 trillion South Korean Won (1.4 billion USD) on data, network, artificial intelligence. Since inauguration, the government has also established The Presidential Committee on the Fourth Industrial Revolution and offered support in data, network, and artificial intelligence, the three innovative industries.

President Moon’s keynote speech not only shared prospects in the artificial intelligence industry at the national level but also confirmed that DEVIEW is the nation’s most prestigious conference in software engineering and that NAVER Corporation is the front runner in the South Korean artificial intelligence market.

Sung Kim (Clova AI Head) is giving the keynote speech in Korean.

Sung Kim (Clova AI Head) introduced various state-of-the-art AI technologies developed by Clova AI and shared the company’s vision and aspiration in becoming the world’s leading AI company. He first started by sharing voice synthesis and recognition technologies that can accurately track each individual’s voice in a meeting and take minutes based on the information gathered. He played a voice AI that cannot only display happiness and sadness perfectly but also generate synthesized voice files with dialect somewhat successfully. He then moved onto sharing successful OCR cases, introducing how the OCR team was able to generate Korean handwriting fonts from 256 characters only and how the recognition module extracts the necessary information on receipts and bills from images. He showed how the AI call agent accepts a restaurant reservation phone call by playing the demo of AI Call. He stated that Clova AI proactively engages in bringing positive impacts to the AI ecosystem by releasing its API’s to the cloud, holding hackathons and software engineering conference, such as DEVIEW, and publishing papers at the world-renowned conferences.

On a side note, there was a remarkable coincidence. ICCV 2019 was located at the same venue and had an overlapping schedule with Deview: also at COEX Convention Center, October 27 to November 2.

Clova AI’s researchers also delivered lectures on their projects. Please click on the links to the slides to view the lecture slides.

1엄~청 큰 언어 모델 공장 가동기!
(LaRva: Language Representation by Clova)
Dongjun Lee 
SungDong Kim
Clova AISlides
Speech
2나 대신 손글씨 써주는 AI 만들기 
(성공적인 Side Project)
Bado LeeClova AISlides
Speech
3신호처리 이론으로 실용적인 스타일 변환 모델 만들기
(Better Faster Stronger Transfer)
Jaejun YooClova AISlides
Speech
4챗봇 1만 개의 모델 서빙하기:
AI 서비스 어디까지 해봤니
SukHyun KoClova AI BusinessSlides
Speech
5레이블링 조금 잘못돼도 괜찮아요: Clova가 레이블 노이즈 잡는 법Jaewook KangClova AI BusinessSlides
Speech
6어디까지 깎아봤니?: 
모바일 서비스를 위한 가벼운 이미지 인식/검출 딥러닝 모델 설계
Dongyoon HanClova AI ResearchSlides
Speech
7예약 전화도 쉽게 받는 인공지능 비서를 만드는 P;ㅠ 
(피.땀.눈물)
Kyoung Tae DoClova AISlides
Speech
  1. 엄~청 큰 언어 모델 공장 가동기! (LaRva: Language Representation by Clova)
    A pipeline that is optimized for training and evaluation of large-scale models, LaRva has testified its high performance, winning the first place on the KorQuAD v1.0 (Korean Q & A dataset) leaderboard that was hosted by LG CNS. The pre-training process of a large language model needs a very large resource and a significant amount of time. DongJun Lee and Sungdong Kim talk about several attempts to lessen the size of the resource and the amount of time. Also, the speakers introduce the process of building an optimized pipeline and share the relevant experimental results. They share how they have tried to apply the big model to diverse NLP services and where their research will be directed.
  2. 나 대신 손글씨 써주는 AI 만들기 (성공적인 Side Project)
    To improve the OCR handwriting recognition accuracy, Bado Lee strived to improve Korean handwritten font generation technology to the applicable level. In this speech, he shares how he was able to achieve the mission. The speaker introduces the state-of-the-art font generation technologies, success in and failure of applying such technologies to the field of handwriting font generation, the developed handwritten font generation method, and required data. 
  3. 신호처리 이론으로 실용적인 스타일 변환 모델 만들기 (Better Faster Stronger Transfer)
    In his speech, Jaejun Yoo introduces his research on a faster, better, and stronger photorealistic style transfer model, WCT2, that was also presented at ICCV 2019. Jaejun Yoo was inspired by the signal processing theory and invented the WCT2 model, a smaller-than-the-existing-models produces previously impossible, high-resolution images (1024×1024) in only four seconds. WCT2 is the first end-to-end photorealistic style transfer model with a lighter conversion. The speaker explains the carefully selected relevant studies that played an essential role in the development process of WCT2, the development process, motivation, and the limitations of each technology.
  4. 챗봇 1만 개의 모델 서빙하기: AI 서비스 어디까지 해봤니
    In this session, Sukhyun Ko shares his experience in the development and launch of the chatbot engine-builder, including episodes in learning and batching of over 100,000 models and serving a customer-specific model on a large scale. The speaker talks about how he created large data from a tool, saved them, managed a snapshot, built an organic system that learns through a learning cluster, and served tens of thousands of models. He discusses the data processing methodology in the AutoML era. He also views his models in two different viewpoints: from the engineering perspective, the optimization of the model’s serving environment, and, from the business perspective, concerns in creating a marketable service.
  5. 레이블링 조금 잘못돼도 괜찮아요: Clova가 레이블 노이즈 잡는 법
    “Garbage in garbage out!”
    Given that the data quality is poor, an individual cannot train good models. Therefore, it is essential to find universal ways to improve data quality for AI researchers. The lecture discusses the AutoML methodology to improve the quality of the labeled datasets. First, Jaewook Kang talks about a label noise that affects the data quality and introduces the PICO (Probabilistic Iterative Correction) algorithm, explaining the theoretical background of how PICO can eliminate label noises and intuition and principles of the algorithm. He concludes the lecture by introducing application examples of PICO to several FAQ service domains.
  6. 어디까지 깎아봤니?: 모바일 서비스를 위한 가벼운 이미지 인식/검출 딥러닝 모델 설계
    Object recognition, detection, and segmentation, which are essential for most AI-related services using images, require a high-performance backbone model. Dongyoon Han introduces reXNet, a new lightweight backbone model developed by Clova AI. First, he shares several important tips for training existing backbone models, including ResNet. Also, he then shares finetuning results from reXNet that were applied to image detection, OCR detector, etc. Now embedded and deployed in Clova’s OCR technology, this model is currently under finetuning tests for each task to be applied to other services soon.
    Further details of reXNet will be released on ArXiv in December.
  7. 예약 전화도 쉽게 받는 인공지능 비서를 만드는 P;ㅠ (피.땀.눈물)
    This project has started by asking the following question: “who will make a restaurant reservation on the phone these days?” Kyung Tae Do has found out in this project that the previously-set assumptions were all wrong and that there are still a lot of problems in the world that is left for AI. How is making a phone assistant different from making a speaker assistant? Kyung Tae Do shares experience in making and installing a phone assistant in many restaurants, including Outback Steakhouse Migeum. The speaker presents solutions towards many difficulties in engineering, deep learning model, and UX.

Also, see the following link to check the full lecture schedule and download the lecture notes.

https://deview.kr/2019/schedule

Media Coverage of DEVIEW 2019 (in English)