DEEP JOINT TRANSMISSION-RECOGNITION FOR POWER-CONSTRAINED IOT DEVICES

In this paper, the deep source channel joint coding is proposed （JSCC） Combine with network pruning , And apply it to image classification on wireless network . The main innovation is in JSCC Network pruning is introduced on .

The main problem to be solved ：

 1、minimizing the number of computations that have to be done locally;

2、designing a robust communication scheme within the limited available transmission power.

Put forward a four step training strategy ：

1、 Preliminary training VGGnet

2、 application “Pruning convolutional neural networks for resource efficient inference,” The pruning algorithm in this paper prunes the network of devices .

3、 Use each image in the training set to run a pruning part of the forward traversal network , To extract all possible feature maps at the segmentation points . Use the feature graph as a new training set of our automatic encoder , And train the automatic encoder .

4、 Micro training of the whole network .

experimental result ：

The future work ：

1、 To further develop this design , It is planned to reduce the number of training steps by merging and pruning to the joint training stage .

2、 Design communication awareness pruning , It can not only reduce the computational complexity , And its filter significance measurement is directly based on wireless channel conditions