Co-evolution of Novel Tree-Like ANNs and Activation Functions: An Observational Study
Lists
Only you see this
Abstract
Deep convolutional neural networks (CNNs) represent the state-of-the-art model structure in image classification problems. However, deep CNNs suffer from issues of interpretability and are difficult to train. This work presents new tree-like shallow ANNs, and offers a novel approach to exploring and examining the relationship between activation functions and network performance. The proposed work is examined on the MNIST and CIFAR10 datasets, finding surprising results relating to the necessity and benefit of activation functions in this new type of shallow network. In particular the work finds high accuracy networks for the MNIST dataset which utilise pooling operations as the only non-linearity, and demonstrate a certain invariance to the specific form of activation functions on the more complicated CIFAR10 dataset.
BibTeX
@inproceedings{oneill2018co,
author = {Damien O'Neill and Bing Xue and Mengjie Zhang},
title = {{Co-evolution of Novel Tree-Like ANNs and Activation Functions: An Observational Study}},
booktitle = {Australasian Conference on Artificial Intelligence},
year = {2018},
url = {https://openreview.net/forum?id=sFHzNWIFeH},
}Loading related content...