Abstract
An auto-encoder which can be split into two parts is designed. 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. It is
tested by tensorflow and mnist dataset. The abstract network is like
LeNet-5. The concrete network is the inverse of the abstract network. A
picture can change to label that is compression, then change it back
from label that is decompression. So compression and decompression can
be realized by the autoencoder. Through test, the absolute function can
do generation task well, while the leaky relu function can do
classification task well. Lossy compression can be achieved by abstract
network and concrete network with absolute function. With one-hot
encoding, the compression ratio is 5.1% when decompression quality is
good. With binary encoding, the compression ratio is 2% when
decompression quality is ok. When jump connection and negative feedback
are used, the decompression performance is good.