Architecture Design for Convolutional Auto-Encoders

Abstract

AbstractAlthough the CAE and its variations have proven benefits in a variety of applications, one key restriction is that their stacked architectures are incompatible with those of state-of-the-art CNNs. The amount of convolutional and pooling layers in the stacked CNN is the same because each CAE contains a pooling layer and a convolutional layer in the encoder. State-of-the-art CNNs, on the other hand, have different amounts of convolutional and pooling layers. The constraint on the number of convolutional and pooling layers contained in CAEs ought to be lifted because the architecture of CNN is one of the important ingredients contributing to the final performance. However, because of the non-differentiable and non-convex properties in practice, it is intractable to determine optimal numbers for convolutional layers and pooling layers, which is related to NAS.