Fabio A. Arciniegas

Neural networks for text-to-speech phoneme recognition

Presents two different artificial neural network (ANN) approaches for phoneme recognition for text-to-speech applications: staged backpropagation neural networks and self-organizing maps. Several current commercial approaches rely on an exhaustive dictionary approach for text-to-phoneme conversion. Applying neural networks to phoneme mapping for text-to-speech conversion creates a fast distributed recognition engine. This engine not only supports the mapping of missing words in the database, but it can also mitigate contradictions related to different pronunciations for the same word. The ANNs presented in this work were trained based on the 2,000 most common words in American English. Performance metrics for the 5,000, 7,000 and 10,000 most common words in English were also estimated to test the robustness of these neural networks.

SVM sensitivity analysis: an application to currency crises aftermaths

A currency crisis is an economic event where a countrys fixed exchange rate is under pressure by speculators. In some cases, currency crises are followed by strong recessions (e.g., recent Asian and Argentinean crises), but in other cases they are not. This paper seeks to determine what are the most significant factors in explaining the consequences of currency crises on the economy. This paper collects data on 25 variables for 64 currency crises between 1970 and 1999. This research uses a novel algorithm with support vector machines (SVM) for selecting significant variables. This algorithm works well with datasets characterized by nonlinearity and low variable-observation ratio. Variables of banking size and fragility, international trade, and devaluation were the most significant. Variables of banking supervision, economic development, and IMF intervention were found less significant. The variable selection results of the algorithm were compared with all-best subsets variable selection. The results of our algorithm are more consistent with the economic literature than the results from all-best subsets.

Bagging neural network sensitivity analysis for feature reduction for in-silico drug design

This paper illustrates a new approach to sensitivity analysis for feature selection using multiple ensemble neural networks in a bootstrapping mode with bagging. This methodology is applied to in-silico drug design with QSAR (quantitative structural activity relationship), which is notoriously challenging for machine learning because typically there are on the order of 300-1000 dependent features, often for as few as 50-100 data points. For an HIV dataset with 160 wavelets descriptors, the number of relevant features was reduced to 35, and the resulting predictive neural network model gave better results than with the full feature set.

Data Mining for Molecules with 2-D Neural Network Sensitivity Analysis

This paper illustrates a data mining application using two-dimensional (2-D) neural network sensitivity analysis for gaining insight into data strip mining problems. Data strip mining refers to predictive data mining problems where there are a large number of descriptive features, and the number of features is on the order of or exceeds the number of data records (e.g., 100 to 1000 features for 50 to 300 data records). After reducing the number of descriptive features to a manageable set using 1-D neural network sensitivity analysis (e.g., 40 features), a 2-D neural network sensitivity analysis allows the user to visualize variations in the response to identify relevant combinations of features. Each relevant combination can then be analyzed independently to look for interesting patterns and relationships, and can be used in this way to either prune more features or to get insight into the underlying rules for the model. 2-D sensitivity analysis enables the exploration of relevant relationships and features resulting in more robust, meaningful, and efficient models. This methodology was applied to an in-silico drug design problem with 64 molecules and 160 d escriptive features.