Convolutional auto-encoders have shown their remarkable performance instacking to deep convolutional neural networks for classifying image dataduring past several years. However, they are unable to construct thestate-of-the-art convolutional neural networks due to their intrinsicarchitectures. In this regard, we propose a flexible convolutional auto-encoderby eliminating the constraints on the numbers of convolutional layers andpooling layers from the traditional convolutional auto-encoder. We also designan architecture discovery method by using particle swarm optimization, which iscapable of automatically searching for the optimal architectures of theproposed flexible convolutional auto-encoder with much less computationalresource and without any manual intervention. We use the designed architectureoptimization algorithm to test the proposed flexible convolutional auto-encoderthrough utilizing one graphic processing unit card on four extensively usedimage classification datasets. Experimental results show that our work in thispaper significantly outperform the peer competitors including thestate-of-the-art algorithm.
展开▼
