We present an original initialization procedure for the parameters of feedforward wavelet networks, prior to training by gradient-based techniques. It takes advantage of wavelet frames stemming from the discrete wavelet transform, and uses a selection method to determine a set of best wavelets whose centers and dilation parameters are used as initial values for subsequent training. Results obtained for the modeling of two simulated processes are compared to those obtained with a heuristic initialization procedure, and demonstrate the effectiveness of the proposed method.

In the framework of nonlinear process modeling, we propose training algorithms for feedback wavelet networks used as nonlinear dynamic models. An original initialization procedure is presented, that takes the locality of the wavelet functions into account. Results obtained for the modeling of several processes are presented; a comparison with networks of neurons with sigmoidal functions is performed.

The article compares two approaches to the description of ultrasound vocal tract images for application in a "silent speech interface," one based on tongue contour modeling, and a second, global coding approach in which images are projected onto a feature space of Eigentongues. A curvaturebased lip profile feature extraction method is also presented. Extracted visual features are input to a neural network which learns the relation between the vocal tract configuration and line spectrum frequencies (LSF) contained in a one-hour speech corpus. An examination of the quality of LSF's derived from the two approaches demonstrates that the eigentongues approach has a more efficient implementation and provides superior results based on a normalized mean squared error criterion.Index Terms-image processing, speech synthesis, neural network applications, communication systems, silent speech interface

No abstract

Abstract.A statistical method trained and optimized to retrieve seven-layer relative humidity (RH) profiles is presented and evaluated with measurements from radiosondes. The method makes use of the microwave payload of the Megha-Tropiques platform, namely the SAPHIR sounder and the MADRAS imager. The approach, based on a generalized additive model (GAM), embeds both the physical and statistical characteristics of the inverse problem in the training phase, and no explicit thermodynamical constraint -such as a temperature profile or an integrated water vapor content -is provided to the model at the stage of retrieval. The model is built for cloud-free conditions in order to avoid the cases of scattering of the microwave radiation in the 18.7-183.31 GHz range covered by the payload. Two instrumental configurations are tested: a SAPHIR-MADRAS scheme and a SAPHIR-only scheme to deal with the stop of data acquisition of MADRAS in January 2013 for technical reasons. A comparison to learning machine algorithms (artificial neural network and support-vector machine) shows equivalent performance over a large realistic set, promising low errors (biases < 2.2 %RH) and scatters (correlations > 0.8) throughout the troposphere (150-900 hPa). A comparison to radiosonde measurements performed during the international field experiment CINDY/DYNAMO/AMIE (winter 2011-2012) confirms these results for the mid-tropospheric layers (correlations between 0.6 and 0.92), with an expected degradation of the quality of the estimates at the surface and top layers. Finally a rapid insight of the estimated large-scale RH field from Megha-Tropiques is presented and compared to ERA-Interim.