“Deep learning” methods, such as deep convolutional networks, have inspired a renaissance in neural network use, and are becoming the default option for difficult problems like image and speech recognition. However, here we show that state-of-the-art results for some well-known image classification benchmarks can be surpassed using a “shallow” convolutional neural network. The method relies on the presence of untrained randomly-valued input weight layers, and tuning only of output weights, resulting in a convex objective function that is optimised in one shot using the entire training dataset. In combination with its competitive classification performance, the method’s rapid training speed and low number of tunable parameters suggest strong potential as an alternative to deep-learning methods when used in embedded hardware applications requiring frequent retraining or online training.
Assoc. Prof. Mark McDonnell received a PhD in electronic engineering in 2006, from The University of Adelaide, Australia. He is currently Associate Research Professor in Computational Neuroscience at the University of South Australia, which he joined in 2007. Within the University of South Australia, McDonnell currently holds a five-year ARC Australian Research Fellowship, and is Principal Investigator and Founder of the Computational and Theoretical Neuroscience Laboratory. McDonnell has published over 80 papers, including several review articles, a patent, and a book on stochastic resonance, published by Cambridge University Press. McDonnell is a member of the editorial boards of PLOS One and Fluctuation and Noise Letters, and has served as Chief Guest Editor for Proceedings of the IEEE and Frontiers in Computational Neuroscience. He is a Senior Member of the IEEE and a Founding Member of Neuroeng: The Australian Association of Computational Neuroscientists and Neuroengineers.
Computational and Theoretical Neuroscience Laboratory
Institute for Telecommunications Research
School of Information Technology and Mathematical Sciences
University of South Australia
|Thursday, 2nd April, 2015
LG28 Napier Building (Lower Ground)