Deploying deep neural networks for edge computer vision (CV) applications is a challenging task. Edge devices have limited capacity, requiring computationally efficient models. Edge AI applications are also diverse, requiring different recognition capabilities for different tasks. Following the traditional supervised learning paradigm, each task requires collecting task-specific data, and then re-designing and training task-specific neural networks. This is not scalable when facing wide varieties of edge CV applications. This talk introduces some recent efforts aiming to address the above two challenges. First, we discuss how we built the latest generation of efficient CV models. Second, we discuss how we use natural language as supervision to train neural networks to recognize new concepts in a zero-shot manner, without task-specific data.
Want to sponsor? Contact us to find out more.