Speaker
Daniel Kuhse
Description
Anytime algorithms can be interrupted before completion while still delivering an intermediate result. This is a desirable property for embedded systems where timing is critical, such as object detection in cyber-physical-systems. However it is generally neither supported by models nor inference frameworks. To enable a model to be anytime, early-exits can be added to the network, which allow an earlier branching off from intermediate layers. We present initial explorations on how to formally define anytime quality, modify the well-known YOLO object detection model to be anytime, in regards to how early-exits can be added and how the architecture can be modified, and the current feasibility of anytime inference.
