Class-specific Models Research Articles

Recognizing handwritten digits using hierarchical products of experts

The product of experts learning procedure can discover a set of stochastic binary features that constitute a nonlinear generative model of handwritten images of digits. The quality of generative models learned in this way can be assessed by learning a separate model for each class of digit and then comparing the unnormalized probabilities of test images under the 10 different class-specific models. To improve discriminative performance, a hierarchy of separate models can be learned, for each digit class. Each model in the hierarchy learns a layer of binary feature detectors that model the probability distribution of vectors of activity of feature detectors in the layer below. The models in the hierarchy are trained sequentially and each model uses a layer of binary feature detectors to learn a generative model of the patterns of feature activities in the preceding layer. After training, each layer of feature detectors produces a separate, unnormalized log probability score. With three layers of feature detectors for each of the 10 digit classes, a test image produces 30 scores which can be used as inputs to a supervised, logistic classification network that is trained on separate data.

An Analysis of Relationships among Climate Forcing and Time-Integrated NDVI of Grasslands over the U.S. Northern and Central Great Plains

Time-integrated normalized difference vegetation index (TI NDVI) derived from the multitemporal satellite imagery (1989–1993) was used as a surrogate for primary production to investigate climate impacts on grassland performance for central and northern Great Plains grasslands. Results suggest that spatial and temporal variability in growing season precipitation, potential evapotranspiration, and growing degree days are the most important controls on grassland performance and productivity. When TI NDVI and climate data of all grassland land cover classes were examined as a whole, a statistical model showed significant positive correlation between the TI NDVI and accumulated spring and summer precipitation, and a negative correlation between TI NDVI and spring potential evapotranspiration. The coefficient of determination (R 2) of the general model was 0.45. When the TI NDVI-climate relationship was examined by individual land cover type, the relationship was generally better defined in terms of the variance accounted for by class-specific models R 2=0.39–0.94 . The photosynthetic pathway is an important determinant of grassland performance with northern mixed prairie (mixture of C 3 and C 4 grassland) TI NDVI affected by both thermal and moisture conditions during the growing season while southern plains grasslands (primarily C 4 grassland) were predominantly influenced by spring and summer precipitation. Grassland land cover classes associated with sandy soils also demonstrated a strong relationship between TI NDVI and growing season rainfall. Significant impact of interannual climate variability on the TI NDVI–climate relationship was also observed. The study suggests an integrated approach involving numerical models, satellite remote sensing, and field observations to monitor grassland ecosystem dynamics on a regional scale.

Model-based covariance mean variance classification techniques: algorithm development and application to the acoustic classification of zooplankton

For inversion problems in which the theoretical relationship between observed data and model parameters is well characterized, a promising approach to the classification problem is the application of techniques that capitalize on the predictive power of class-specific models. Theoretical models have been developed for three zooplankton scattering classes (hard elastic-shelled, e.g., pteropods; fluid-like, e.g., euphausiids; and gas-bearing, e.g., siphonophores), providing a sound basis for model-based classification approaches. The covariance mean variance classification (CMVC) techniques classify broad-band echoes from individual zooplankton based on comparisons of observed echo spectra to model space realizations. Three different CMVC algorithms were developed: the integrated score classifier, the pairwise score classifier, and the Bayesian probability classifier; these classifiers assign observations to a class based on similarities in covariance, mean, and variance while accounting for model spare ambiguity and validity. The CMVC techniques were applied to broad-band (/spl sim/350-750 kHz) echoes acquired from 24 different zooplankton to invert for scatterer class and properties. All three classification algorithms had a high rate of success with high-quality high SNR data. Accurate acoustic classification of zooplankton species has the potential to significantly improve estimates of zooplankton biomass made from ocean acoustic backscatter measurements.

Adsorption from aqueous phase by activated carbon: a simplified application of the solvophobic theory

A simple approach to predict adsorption capacity of activated carbon is presented. Developed from the solvophobic theory of adsorption, this simplified approach uses molecular descriptors that are easy to calculate from the molecular structure of the adsorbates. A distinctive feature of this approach is that a single generalized model could predict the adsorbability of different classes of compounds. The predictive abilities of this general model and the different class-specific models derived by using the regular solvophobic approach are compared to demonstrate better quality of prediction possible by this general model.