상세정보
학술대회
Deployment Framework Design Techniques for Optimized Neural Network Applications
- 발행일
- 202210
- 출처
- International Conference on Information and Communication Technology Convergence (ICTC) 2022, pp.2312-2314
- 협약과제
- 22HS2800, 신경망 응용 자동생성 및 실행환경 최적화 배포를 지원하는 통합개발 프레임워크 기술개발, 조창식
- 초록
- For companies with limited neural network-based technology, rapid and accurate development of AI services requires an integrated development framework that facilitates a device adaptive neural network model search and a target environment optimization code generation and deployment. For the efficient execution of neural network applications, it is necessary to have a technique capable of solving complex and difficult problems such as iterative application of various optimization techniques, verification of errors, and analysis of causes of performance degradation. This paper proposes the design of DeepFramework that can deploy neural network application services. DeepFramework generates neural networks according to user requirements. Next, the neural network generated according to the user requirements specification is optimized, and application templates are automatically generated and deployed. DeepFramework is currently under development, with plans to release later this year. Additionally, we experimented with vision recognition neural network applications for validation of DeepFramework. In the future, DeepFramework will open the results after excluding the technology dependency of a specific company. This makes it possible to develop neural network application services quickly and easily at low cost in various industries such as business, medical care, and smart factories.
- KSP 제안 키워드
- Adaptive neural network, Design techniques, Development framework, Environment optimization, Framework Design, Low-cost, Neural network applications, Optimization code, Optimization techniques(OT), Optimized neural network, Smart Factory