Equivalence of backpropagation and contrastive Hebbian learning in a layered network

Backpropagation and contrastive Hebbian learning are two methods of training networks with hidden neurons. Backpropagation computes an error signal for the output neurons and spreads it over the hidden neurons. Contrastive Hebbian learning involves clamping the output neurons at desired values and letting the effect spread through feedback connections over the entire network. To investigate the relationship between these two forms of learning, we consider a special case in which they are identical: a multilayer perceptron with linear output units, to which weak feedback connections have been added. In this case, the change in network state caused by clamping the output neurons turns out to be the same as the error signal spread by backpropagation, except for a scalar prefactor. This suggests that the functionality of backpropagation can be realized alternatively by a Hebbian-type learning algorithm, which is suitable for implementation in biological networks.     

Electron-beam treatment of aromatic hydrocarbons that can be air-stripped from contaminated groundwater. 2. Gas-phase studies

The electron-beam (EB) degradation of volatile aromatics (benzene, toluene, ethylbenzene, xylenes: BTEX) in groundwater strip gas, which in the present work has been modeled by the introduction of the desired aromatic(s) to a stream of air or another gas, such as oxygen, is initiated essentially by the addition of *OH radicals to the aromatic ring, giving rise to hydroxycyclohexadienyl radicals, which form the corresponding peroxyl radicals upon addition of oxygen. As studied in some detail with benzene as a BTEX representative, various reactions of these lead to numerous oxidation products in a cascade of reactions, including the decomposition of products under the prevailing conditions of high turnover of the initial aromatic. Importantly, hydroxycyclohexadienylperoxyl radical formation is partly reversible, and the reactions of the hydroxycyclohexadienyl radicals, which thus have a significant presence in these systems, must therefore also be taken into consideration. In the gas phase, in contrast to the aqueous phase (see Part 1), the reactions of the hydroxycyclohexadienyl radicals lead to oligomeric products that appear to contribute, in addition to ionic clusters, to nucleation for the aerosols observed. Various nitrated products, among them nitrophenols, are observed when air is used for the stripping. However, these studies did not clear the pilot plant stage, since BTEX degradation using a bioreactor carried out in parallel was so successful that the EB technology was judged to be noncompetitive. As for the latter, expensive equipment consisting of a stripper, the EB machine, and an aerosol precipitator would be required. The condensed aerosols are biorefractory and would require further treatment for detoxification.     

A comparison of pebble tree transducers with macro tree transducers

The n-pebble tree transducer was recently proposed as a model for XML query languages. The four main results on deterministic transducers are: First, (1) the translation of an n-pebble tree transducer can be realized by a composition of n+1 0-pebble tree transducers. Next, the pebble tree transducer is compared with the macro tree transducer, a well-known model for syntax-directed semantics, with decidable type checking. The -pebble tree transducer can be simulated by the macro tree transducer, which, by the first result, implies that (2) can be realized by an (n+1)-fold composition of macro tree transducers. Conversely, every macro tree transducer can be simulated by a composition of 0-pebble tree transducers. Together these simulations prove that (3) the composition closure of n-pebble tree transducers equals that of macro tree transducers (and that of 0-pebble tree transducers). Similar results hold in the nondeterministic case. Finally, (4) the output languages of deterministic n-pebble tree transducers form a hierarchy with respect to the number n of pebbles.This revised version was published online in September 2003 with corrections to type sizes.Received: 16 January 2003, Sebastian Maneth: Present address: Swiss Institute of Technology Lausanne, Programming Methods Laboratory (LAMP), 1015 Lausanne, Switzerland, e-mail (sebastian.maneth@epfl.ch)     

Boundary layer scattering measurements with a charge-coupled device camera lidar

A CCD-based bistatic lidar (CLidar) system has been developed and constructed to measure scattering in the atmospheric boundary layer. The system uses a CCD camera, wide-angle optics, and a laser. Imaging a vertical laser beam from the side allows high-altitude resolution in the boundary layer all the way to the ground. The dynamic range needed for the molecular signal is several orders of magnitude in the standard monostatic method, but only approximately 1 order of magnitude with the CLidar method. Other advantages of the Clidar method include low cost and simplicity. Observations at Mauna Loa Observatory, Hawaii, show excellent agreement with the modeled molecular-scattering signal. The scattering depends on angle (altitude) and the polarization plane of the laser.     

Recognition of emergent human behaviour in a smart home: A data mining approach

Motivated by a growing need for intelligent housing to accommodate ageing populations, we propose a novel application of intertransaction association rule (IAR) mining to detect anomalous behaviour in smart home occupants. An efficient mining algorithm that avoids the candidate generation bottleneck limiting the application of current IAR mining algorithms on smart home data sets is detailed. An original visual interface for the exploration of new and changing behaviours distilled from discovered patterns using a new process for finding emergent rules is presented. Finally, we discuss our observations on the emergent behaviours detected in the homes of two real world subjects.     

Experimental and Numerical Failure Criteria for Sheet Metal Forming

For sheet metal forming, often the forming limit diagram (FLD) is used as failure criterion as it can be derived easily in experiments. It is based on the assumption that localization of strain in the sheet plane is responsible for crack initiation, but application of FLD is limited to linear strain paths. Hence, only forming processes with approximately the same deformation history as the experiments carried out for FLD determination should be evaluated by this criterion. Forming limit stress diagrams (FLSD) do not exhibit such strict limitations. They are based on the assumption that principal stresses in the sheet plane are responsible for crack initiation. As these stresses are usually calculated by FE analysis using elastic plastic material laws, strain hardening is considered. Two‐step forming tests as application examples prove the FLSD to be adequate for evaluation of non‐linear forming processes with alternating forming directions. Nevertheless, FLSD are derived in extensive investigations which makes them unattractive for most industrial applications. Furthermore, both FLD and FLSD do not consider the physical background of ductile crack initiation which is provoked by an interaction of local stress triaxiality and equivalent plastic strain. Hence, a reliable failure criterion should concentrate on these two parameters. The Gurson‐Tvergaard‐Needleman‐ (GTN‐) damage model can predict crack initiation during sheet metal forming. Application of the GTN model to 2 step forming tests with the bake hardening steel H220BD+Z showed good agreement to experimental results although a sensitivity of the model to mesh size and stress triaxiality is observed.     

A within-modality test of semantic knowledge: The Size/Weight Attribute Test.

Individuals with semantic dementia (SD; n=12) or Alzheimer's disease (AD; n=20) and healthy volunteers (n=40) were tested on tasks probing attribute rather than associative knowledge of animals and objects in the visual and verbal domains. The tasks were within modality, in that they probed knowledge in a single presentation modality (pictures or written words) and did not require cross-modal matching. Participants were required to make the same simple judgment about triads of animal stimuli ("Which is the largest and smallest?") and object stimuli ("Which is the heaviest and lightest?"). Control participants scored at ceiling on this simple task. Overall, the SD patients were significantly more impaired on this task than were the AD patients, who in turn were significantly more impaired than the controls. There was a strong trend for SD but not AD patients to show worse performance with verbal than visual material. However, no significant effects of category were found in either patient group. The Size/Weight Attribute Test has the potential to assess modality-specific deficits of semantic knowledge. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

Biomimetic Concealing of PLGA Microspheres in a 3D Scaffold to Prevent Macrophage Uptake

Scaffolds functionalized with delivery systems for the release of growth factors is a robust strategy to enhance tissue regeneration. However, after implantation, macrophages infiltrate the scaffold, eventually initiating the degradation and clearance of the delivery systems. Herein, it is hypothesized that fully embedding the poly(d,l ‐lactide‐co‐glycolide acid) microspheres (MS) in a highly structured collagen‐based scaffold (concealing) can prevent their detection, preserving the integrity of the payload. Confocal laser microscopy reveals that non‐embedded MS are easily internalized; when concealed, J774 and bone marrow‐derived macrophages (BMDM) cannot detect them. This is further demonstrated by flow cytometry, as a tenfold decrease is found in the number of MS engulfed by the cells, suggesting that collagen can cloak the MS. This correlates with the amount of nitric oxide and tumor necrosis factor‐α produced by J774 and BMDM in response to the concealed MS, comparable to that found for non‐functionalized collagen scaffolds. Finally, the release kinetics of a reporter protein is preserved in the presence of macrophages, only when MS are concealed. The data provide detailed strategies for fabricating three dimensional (3D) biomimetic scaffolds able to conceal delivery systems and preserve the therapeutic molecules for release.     

FlipCut Supertrees: Towards Matrix Representation Accuracy in Polynomial Time

In computational phylogenetics, supertree methods provide a way to reconstruct larger clades of the Tree of Life. The supertree problem can be formalized in different ways, to cope with contradictory information in the input. In particular, there exist methods based on encoding the input trees in a matrix, and methods based on finding minimum cuts in some graph. Matrix representation methods compute supertrees of superior quality, but the underlying optimization problems are computationally hard. In contrast, graph-based methods have polynomial running time, but supertrees are inferior in quality. In this paper, we present a novel approach for the computation of supertrees called FlipCut supertree. Our method combines the computation of minimum cuts from graph-based methods with a matrix representation method, namely Minimum Flip Supertrees. Here, the input trees are encoded in a 0/1/?-matrix. We present a heuristic to search for a minimum set of 0/1-flips such that the resulting matrix admits a directed perfect phylogeny. We then extend our approach by using edge weights to weight the columns of the 0/1/?-matrix. In our evaluation, we show that our method is extremely swift in practice, and orders of magnitude faster than the runner up. Concerning supertree quality, our method is sometimes on par with the “gold standard” Matrix Representation with Parsimony.