Neural network pruning with Tukey-Kramer multiple comparison procedure

Reducing a neural network's complexity improves the ability of the network to generalize future examples. Like an overfitted regression function, neural networks may miss their target because of the excessive degrees of freedom stored up in unnecessary parameters. Over the past decade, the subject of pruning networks produced nonstatistical algorithms like Skeletonization, Optimal Brain Damage, and Optimal Brain Surgeon as methods to remove connections with the least salience. The method proposed here uses the bootstrap algorithm to estimate the distribution of the model parameter saliences. Statistical multiple comparison procedures are then used to make pruning decisions. We show this method compares well with Optimal Brain Surgeon in terms of ability to prune and the resulting network performance.     

Depletion of Grain-Boundary Phase and Elastic Creep Deformation Upon Ultra-Long Flexural Stress Rupture Experiments of NC-132 Silicon Nitride

A unique grain-boundary structure evolution was observed in two MgO-doped silicon nitride specimens (Norton, NC-132) that were tested in ultra-long flexure stress rupture experiments with an applied stress of 266 MPa and fractured at 14 941 and 17 376 h. Transmission electron microscopy showed that, although the starting material had a secondary glass phase both at multi-grain junctions and along grain boundaries, the tested specimens contained no residual glass phase. Concurrent with the elimination of the secondary glass phase, a continuous network of cracked grain boundaries was observed after long-term flexure testing consistent with the concept of elastic creep. It is, therefore, concluded that at ultra-long annealing times, this material is affected by creep deformation via microcrack nucleation and growth due to the depletion of the amorphous siliceous grain-boundary phase, which is seen as a truly transient, fugitive secondary phase.     

The application of ionizing radiation in reversible addition-fragmentation chain transfer (RAFT) polymerization: Renaissance of a key synthetic and kinetic tool

Ionizing radiation, such as γ, ultraviolet, microwave and X-ray radiation, has long been used in polymer chemistry as a means of initiating polymerization, crosslinking gels and decomposing particular polymer components. More recently, ionizing radiation has found application in tandem with living radical polymerization to form novel polymeric materials with defined molecular weight and narrow molecular weight distribution. In particular, γ-rays and ultraviolet light both have shown promise as sources of initiation in reversible addition-fragmentation chain transfer (RAFT) polymerization. The ability to apply these sources of initiation at low temperatures is useful in applications where elevated temperature is likely to be detrimental to the system, for instance, in preparing protein-polymer conjugates. Similarly, the use of these initiating sources at low temperature is particularly suitable for some monomers, such as allyl compounds, which have not been synthesized using any other living radical approach. The current review examines the development of ionizing radiation as a tool in RAFT polymerization, with particular reference to the elucidation of the polymerization mechanism, the synthesis of high functionality polymers and probing the kinetic parameters of the RAFT process.     

Using the high pressure methane isotherm for determination of pore size distribution of carbon adsorbents

A method of determining pore size distribution, PSD, of carbon adsorbents based on the high pressure methane isotherm is presented. A generic software package, and an IBM compatible PC, have been used to search for a PSD in the form of a histogram. The method relies on a known local isotherm, in this case, assuming a simplified model of infinite slit shaped carbon pores.Three carbons, having very different pore structures: BPL, PX-21, and PVDC, were analyzed using the new method and the results compared with those obtained from subcritical Ar, and N₂ isotherms. The analysis from the high pressure methane isotherm gave results which are different than those from the low pressure low temperature isotherms but not significantly enough to be unrealistic.     

Towards a hydrodynamic modelling framework appropriate for applications in urban flood assessment and mitigation using heterogeneous computing

We present a new hydrodynamic modelling framework capable of fully exploiting modern graphics and central processing units (GPUs and CPUs) from any of the mainstream vendors, to be used in the design and assessment of sustainable drainage systems. A finite-volume Godunov-type scheme is combined with the HLLC Riemann solver to create a robust numerical model which correctly addresses wetting and drying and transient flow conditions, and is suitable for application to a wide range of flood simulations. The software is tested with a three day flood event in Carlisle during 2005, at resolutions from 25 m to 2 m. Run-times are significantly reduced without compromising numerical accuracy. Excellent agreement is found between the simulation results and a comprehensive post-event survey. Changes in sensitivity to Manning's n are examined at different resolutions, with changes to the floodplain found to have little influence at 2 m resolution.