Towards interpretable deep learning with limited examples

通过有限的例子实现可解释的深度学习

• 批准号: DP180100106
• 负责人: Prof Ivor Tsang
• 金额: $ 27.49万
• 依托单位: University of Technology Sydney
• 依托单位国家: 澳大利亚
• 项目类别: Discovery Projects
• 财政年份: 2018
• 资助国家: 澳大利亚
• 起止时间: 2018-01-01 至 2021-12-31
• 项目状态: 已结题
• 来源: https://dataportal.arc.gov.au/RGS/Web/Grants/DP180100106
• 关键词: Towards; interpretable; deep; learning; limited

Existing visual concept detection systems are incapable of detecting ever-evolving concepts in daily life. This project aims to extract patterns that describe the semantics of visual concepts and to develop or adapt knowledge transfer learning technologies for new concepts with limited examples. The expected outcomes will provide major technological breakthroughs for building efficient and interpretable learning systems for visual analysis and will open an entirely new research direction: interpretable deep learning with communication mechanism. This new field and its technologies will help us to recognise misuse of home patient medical devices and unauthorised activity, and enable us to devise effective responses to prevent cyberattacks.

现有的视觉概念检测系统无法检测日常生活中不断发展的概念。该项目旨在提取描述视觉概念语义的模式，并开发或调整知识转移学习技术，用于有限示例的新概念。预期成果将为构建高效、可解释的视觉分析学习系统提供重大技术突破，并将开辟一个全新的研究方向：具有通信机制的可解释深度学习。这一新领域及其技术将帮助我们识别家庭患者医疗设备的滥用和未经授权的活动，并使我们能够制定有效的应对措施来防止网络攻击。