Testing Deep Learning Techniques on Environmental Datasets

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 7 January 2015: 10:30 AM
124B (Phoenix Convention Center - West and North Buildings)
William W. Hsieh, Univ. of British Columbia, Vancouver, BC, Canada

Machine learning has advanced through neural networks (NN) and kernel methods to deep learning methods. Deep learning (DL) has led to the revamping of the traditional multi-layer perceptron NN model -- e.g. using the rectified linear activation function instead of a sigmoidal-type function and the random "dropout" of hidden neurons (and/or input neurons) during training for regularization -- thereby permitting the use of more hidden layers and greater complexity than the traditional NN models. Most AI problems involve classification rather than regression, and have vast amounts of training data, in contrast to most environmental datasets. It is therefore not clear whether the DL techniques are necessarily applicable to environmental datasets. A variety of environmental datasets (e.g. precipitation, air quality, wind power etc.), have been tested using DL techniques for regression to determine under what circumstances are DL techniques better than traditional NN models.