Learning the Structure of Deep Sparse Graphical Models

Deep belief networks are a powerful way to model complex probability distributions. However, learning the structure of a belief network, particularly one with hidden units, is difficult. The Indian buffet process has been used as a nonparametric Bayesian prior on the directed structure of a belief network with a single infinitely wide hidden layer. In this paper, we introduce the cascading Indian buffet process (CIBP), which provides a nonparametric prior on the structure of a layered, directed belief network that is unbounded in both depth and width, yet allows tractable inference. We use the CIBP prior with the nonlinear Gaussian belief network so each unit can additionally vary its behavior between discrete and continuous representations. We provide Markov chain Monte Carlo algorithms for inference in these belief networks and explore the structures learned on several image data sets.

ANNEALING BEHAVIOUR OF HIGH DOSE As** plus IMPLANTS.

The annealing behaviour of doses up to 4. 10**1**6 ions/cm**2 implanted at ion currents up to 10ma is described. Differences between rapid isothermal and furnace annealing in the measured sheet resistances are due to different amounts of diffusion and to loss of dopant by evaporation. Implantation at high currents (10ma) does not appear to affect the quality of the regrown material provided the temperature rise during implantation is small.

Correlation between density of states distributions and the fermi level position in interfacial regions of amorphous silicon thin film transistors

This paper investigates the variation of the integrated density of states with conduction activation energy in hydrogenated amorphous silicon thin film transistors. Results are given for two different gate insulator layers, PECVD silicon oxide and thermally grown silicon dioxide. The different gate insulators produce transistors with very different initial transfer characteristics, but the variation of integrated density of states with conduction activation energy is shown to be similar.

Velocity distributions for low pressure turbines

A parametric set of velocity distributions has been investigated using a flat plate experiment. Three different diffusion factors and peak velocity locations were tested. These were designed to mimic the suction surfaces of Low Pressure (LP) turbine blades. Unsteady wakes, inherent in real turbomachinery flows, were generated using a moving bar mechanism. A turbulence grid generated a freestream turbulence level that is believed to be typical of LP turbines. Measurements were taken across a Reynolds number range of 50,000-220,000 at three reduced frequencies (0.314, 0.628, 0.942). Boundary layer traverses were performed at the nominal trailing edge using a Laser Doppler Anemometry system and hot-films were used to examine the boundary layer behaviour along the surface. For every velocity distribution tested, the boundary layer separated in the diffusing flow downstream of the peak velocity. The loss production is dominated by the mixing in the reattachment process, mixing in the turbulent boundary layer downstream of reattachment and the effects of the unsteady interaction between the wakes and the boundary layer. A sensitive balance governs the optimal location of peak velocity on the surface. Moving the velocity peak forwards on the blade was found to be increasingly beneficial when bubblegenerated losses are high, i.e. at low Reynolds number, at low reduced frequency and at high levels of diffusion. Copyright © 2008 by ASME.