Collaborative Intelligence

人工智能（AI）正在从研究实验室走向真实的世界。将人工智能“带到边缘”的最有希望的途径之一是协作智能（CI），这是一种在边缘设备和云之间共享人工智能推理的框架。在CI中，AI模型的前端通常部署在边缘设备上，在那里执行初始处理和特征计算。然后，这些中间特征被发送到云端，AI模型的后端在那里完成推理。与更典型的基于云或完全基于边缘的AI模型部署相比，CI已被证明具有节省能源和延迟的潜力，但它也带来了新的挑战，需要开发新的科学和工程原理以实现最佳设计。在CI中，在AI模型的信息路径中插入一个容量有限的通道。这需要压缩在边缘子模型处计算的特征。此外，由于通道缺陷而引入特征的错误需要在云端处理，以便执行成功的推理。最后，需要解决与传输数据的隐私有关的问题。拟议的研究计划旨在发现CI系统的基本见解和开发实用设计。在理论方面，我们计划使用信息理论工具来研究CI系统中的信息流。具体来说，我们计划扩展的信息瓶颈原则的能力有限的学习模型，以深入了解CI系统的性能限制。在实践方面，我们将专注于以下短期目标：为CI系统开发新的特征压缩技术;研究CI系统的错误恢复能力并开发适当的错误控制方法;研究隐私问题并提出CI系统中隐私友好推理的解决方案;并将上述解决方案扩展到多输入，多任务CI系统。 拟议的研究计划处于对加拿大具有重要战略意义的两个技术领域的交叉点：人工智能和信息通信技术（ICT）。拟议研究的结果将对这两个领域产生重大影响，并将刺激它们交叉点的进一步创新。其结果将是提高远程监控、交通、自动化和制造、智能家居和城市、农业等行业新兴应用的能源效率、响应能力、隐私和安全性，同时为塑造未来行业的新的、尚未想象的应用奠定基础。在该计划中接受培训的HQP将获得技能和知识，使他们成为加拿大及其他地区这些重要领域的科技领导者。

Artificial Intelligence (AI) is moving from research labs to the real world. One of the most promising avenues for bringing AI "to the edge" is Collaborative Intelligence (CI), a framework in which AI inference is shared between the edge devices and the cloud. In CI, typically, the front-end of an AI model is deployed on an edge device, where it performs initial processing and feature computation. These intermediate features are then sent to the cloud, where the back-end of the AI model completes the inference. CI has been shown to have the potential for energy and latency savings compared to the more typical cloud-based or fully edge-based AI model deployment, but it also introduces new challenges, which require new science and engineering principles to be developed in order to achieve optimal designs. In CI, a capacity-limited channel is inserted in the information pathway of an AI model. This necessitates compression of features computed at the edge sub-model. Moreover, errors introduced into features due to channel imperfections would need to be handled at the cloud side in order to perform successful inference. Finally, issues related to the privacy of transmitted data need to be addressed. The proposed research program aims to discover fundamental insights and develop practical designs for CI systems. On the theory side, we plan to use information-theoretic tools to study information flow in CI systems. Specifically, we plan to extend the Information Bottleneck principle to capacity-limited learning models in order to gain insight into performance limits of CI systems. On the practical side, we will focus on the following short-term objectives: develop novel feature compression techniques for CI systems; study error resilience of CI systems and develop appropriate error control methods; study privacy issues and propose solutions for privacy-friendly inference in CI systems; and extend the above solutions to multi-input, multi-task CI systems. The proposed research program stands at the intersection of two technological areas of great strategic importance to Canada: AI and Information and Communication Technologies (ICT). The outcomes of the proposed research will have significant impact on both areas, and will spur further innovation at their intersection. The result will be the improvement of energy efficiency, responsiveness, privacy and security of emerging applications in remote monitoring, transportation, automation and manufacturing, smart homes and cities, agriculture and other industries, while setting the stage for new, yet-to-be-imagined applications that will shape industries of tomorrow. The HQP trained in the program will gain skills and knowledge that will position them to become scientific and technological leaders in these important areas, in Canada and beyond.