University of Wollongong

About: University of Wollongong is a education organization based out in Wollongong, New South Wales, Australia. It is known for research contribution in the topics: Population & Graphene. The organization has 15674 authors who have published 46658 publications receiving 1197471 citations. The organization is also known as: UOW & Wollongong University.

Lessons in neural network training: overfitting may be harder than expected

Abstract: For many reasons, neural networks have become very popular AI machine learning models. Two of the most important aspects of machine learning models are how well the model generalizes to unseen data, and how well the model scales with problem complexity. Using a controlled task with known optimal training error, we investigate the convergence of the backpropagation (BP) algorithm. We find that the optimal solution is typically not found. Furthermore, we observe that networks larger than might be expected can result in lower training and generalization error. This result is supported by another real world example. We further investigate the training behavior by analyzing the weights in trained networks (excess degrees of freedom are seen to do little harm and to aid convergence), and contrasting the interpolation characteristics of multi-layer perceptron neural networks (MLPs) and polynomial models. (overfitting behavior is very different - the MLP is often biased towards smoother solutions). Finally, we analyze relevant theory outlining the reasons for significant practical differences. These results bring into question common beliefs about neural network training regarding convergence and optimal network size, suggest alternate guidelines for practical use (lower fear of excess degrees of freedom), and help to direct future work (e.g. methods for creation of more parsimonious solutions, importance of the MLP/BP bias and possibly worse performance of "improved" training algorithms).

Effects of weather, fuel and terrain on fire severity in topographically diverse landscapes of south-eastern Australia

Abstract: The effects of weather, terrain, fuels on fire severity were compared using remote sensing of the severity of two large fires in south-eastern Australian forests. The probability of contrasting levels of fire severity (fire confined to the understorey vs. tree canopies consumed) was analysed using logistic regression. These severities equate to extremes of fire intensity ( 10,000 kW m−1), consequent suppression potential (high vs. nil) and potential adverse ecological impacts on vertebrate fauna and soils (low vs. high). Weather was the major influence on fire severity. Crown fire was absent under non-extreme weather and but more likely under extreme weather, particularly on ridges in vegetation unburnt for >10 years. Crown fire probability was very low in recently burnt vegetation (1–5 years) and increased at higher fuel ages. In all cases, fire severity was lower in valleys, probably due to effects of wind protection and higher fuel moisture in moderating fire behaviour. Under non-extreme weather, fires are likely to be suppressible and burn heterogeneously, due to the influence of topographic position, slope and fuel load. Under extreme weather, fires are influenced only by fuel and topographic position, and probability of suppression on accessible ridges will be low except in recently burnt (i.e. 1–5 year old) fuels. Topographically imposed variation may mitigate adverse ecological effects on arboreal fauna and soil erosion potential.

The Cathode Choice for Commercialization of Sodium-Ion Batteries: Layered Transition Metal Oxides versus Prussian Blue Analogs

Abstract: With the unprecedentedly increasing demand for renewable and clean energy sources, the sodium-ion battery (SIB) is emerging as an alternative or complementary energy storage candidate to the present commercial lithium-ion battery due to the abundance and low cost of sodium resources. Layered transition metal oxides and Prussian blue analogs are reviewed in terms of their commercial potential as cathode materials for SIBs. The recent progress in research on their half cells and full cells for the ultimate application in SIBs are summarized. In addition, their electrochemical performance, suitability for scaling up, cost, and environmental concerns are compared in detail with a brief outlook on future prospects. It is anticipated that this review will inspire further development of layered transition metal oxides and Prussian blue analogs for SIBs, especially for their emerging commercialization.

Dissociation mediates the relationship between childhood trauma and hallucination-proneness.

Abstract: Results. Compared to both healthy and non-hallucinating clinical controls, hallucinating patients reported both significantly higher dissociative tendencies and childhood sexual abuse. Dissociation positively mediated the effect of childhood trauma on hallucination-proneness. This mediational role was particularly robust for sexual abuse over other types of trauma. Signal detection abnormalities were evident in hallucinating patients and patients with a history of hallucinations, but were not associated with pathological dissociative symptoms. Conclusions. These results are consistent with dissociative accounts of the trauma-hallucinations link. Dissociation, however, does not affect reality discrimination. Future research should examine whether other cognitive processes associated with both dissociative states and hallucinations (e.g. deficits in cognitive inhibition) may explain the relationship between dissociation and hallucinatory experiences. Received 16 February 2011 ; Revised 1 July 2011 ; Accepted 13 August 2011 ; First published online 6 September 2011