Neural Architecture Construction using EnvelopeNets
share
In recent years, advances in the design of convolutional neural networks have resulted in significant improvements on the image classification and object detection problems. One of the advances is networks built by stacking complex cells, seen in such networks as InceptionNet and NasNet. These cells are either constructed by hand, generated by generative networks or discovered by search. Unlike conventional networks (where layers consist of a convolution block, sampling and non linear unit), the new cells feature more complex designs consisting of several filters and other operators connected in series and parallel. Recently, several cells have been proposed or generated that are supersets of previously proposed custom or generated cells. Influenced by this, we introduce a network construction method based on EnvelopeNets. An EnvelopeNet is a deep convolutional neural network of stacked EnvelopeCells. EnvelopeCells are supersets (or envelopes) of previously proposed handcrafted and generated cells. We propose a method to construct improved network architectures by restructuring EnvelopeNets. The algorithm restructures an EnvelopeNet by rearranging blocks in the network. It identifies blocks to be restructured using metrics derived from the featuremaps collected during a partial training run of the EnvelopeNet. The method requires less computation resources to generate an architecture than an optimized architecture search over the entire search space of blocks. The restructured networks have higher accuracy on the image classification problem on a representative dataset than both the generating EnvelopeNet and an equivalent arbitrary network.
READ FULL TEXT
