A method for performing online batch normalization, on-device learning, and continuous learning applicable to mobile or IOT devices with further reference to one or more previous batches for use in military purposes, drones or robots. And device, and test method and test device using the same

摘要

A method for online batch normalization, on-device learning, and continuous learning applicable to IOT devices, mobile devices, and the like. A computing device acquires a kth batch with a convolution layer and applies a convolution operation to each input image included in the kth batch to generate a feature map for the kth batch. The loading device has a batch normalization layer, refers to the feature map for the kth batch when k is 1, and is the first previously generated when k is a constant from 2 to m'. To the feature map for the kth batch and the feature map for the kth batch included in at least a part of the previous batch selected from the (k-1)th batch, the adjustment average and the adjustment of the feature map for the kth batch Calculating the variance and applying the batch normalization operation to the feature map for the kth batch. [Selection diagram] Figure 2