A Functionally Separate Autoencoder
<p><b>According to kids’ learning process, an auto</b><b>-</b><b>encoder</b><b> is designed</b><b> which can be split into two parts. The two parts can work well separately.The top half is an abstract network which is trained by supervised learning and can be used to classify and regress. The bottom half is a concrete network which is accomplished by inverse function and trained by self-supervised learning. It can generate the input of abstract network from concept or label. The network can achieve its intended functionality through testing by mnist dataset and convolution neural network.</b><b> R</b><b>ound function</b><b> is added between the abstract network and concrete network in order</b><b> to get the the representative generation of class.</b><b> T</b><b>he generation ability </b><b> can be increased </b><b>by adding jump connection and negative feedback. At last, the characteristics of </b><b>the</b><b> network</b><b> is discussed</b><b>. </b><b>T</b><b>he input can </b><b>be </b><b>change</b><b>d </b><b>to any form by encoder and then change it back by decoder through inverse function. The concrete network can be seen as the memory stored by the parameters.</b><b> </b><b>Lethe is that when new knowledge input,</b><b> </b><b>the training process make</b><b>s</b><b> the parameter</b><b>s</b><b> change.</b><b></b></p>