MAE-NAS: A Low-Cost Customized Model for Object Detection

MAE-NAS is a low-cost customized model that treats deep networks as information systems with continuous state space and maximizes the entropy of the information system. In the process of network modeling, the topological structure of network F is abstracted as a graph G=(V,E), where vertices V represent features and edges E represent various operators. The entropy H(S) measures the total amount of information contained in the entire information system F. Since only the expressive power of the network is concerned in the feature extraction part of object detection, only the information in the vertices is calculated here.