Interpretación geométrica de redes neuronales recurrentes discretas mediante grafos completos

El objetivo del presente trabajo de investigación es explorar nuevas técnicas de implementación, basadas en grafos, para las Redes de Neuronas, con el fin de simplificar y optimizar las arquitecturas y la complejidad computacional de las mismas. Hemos centrado nuestra atención en una clase de Red de Neuronas: las Redes de Neuronas Recursivas (RNR), también conocidas como redes de Hopfield. El problema de obtener la matriz sináptica asociada con una RNR imponiendo un determinado número de vectores como puntos fijos, no está en absoluto resuelto, el número de vectores prototipo que pueden ser almacenados en la red, cuando se utiliza la ley de Hebb, es bastante limitado, la red se satura rápidamente cuando se pretende almacenar nuevos prototipos. La ley de Hebb necesita, por tanto, ser revisada. Algunas aproximaciones dirigidas a solventar dicho problema, han sido ya desarrolladas. Nosotros hemos desarrollado una nueva aproximación en la forma de implementar una RNR en orden a solucionar estos problemas. La matriz sináptica es obtenida mediante la superposición de las componentes de los vectores prototipo, sobre los vértices de un Grafo, lo cual puede ser también interpretado como una coloración de dicho grafo. Cuando el periodo de entrenamiento se termina, la matriz de adyacencia del Grafo Resultante o matriz de pesos, presenta ciertas propiedades por las cuales dichas matrices serán llamadas tetraédricas. La energía asociada a cualquier estado de la red es representado por un punto (a,b) de R2. Cada uno de los puntos de energía asociados a estados que disten lo mismo del vector cero está localizado sobre la misma línea de energía de R2. El espacio de vectores de estado puede, por tanto, clasificarse en n clases correspondientes a cada una de las n diferentes distancias que puede tener cualquier vector al vector cero. La matriz (n x n) de pesos puede reducirse a un n-vector; de esta forma, tanto el tiempo de computación como el espacio de memoria requerido par almacenar los pesos, son simplificados y optimizados. En la etapa de recuperación, es introducido un vector de parámetros R2, éste es utilizado para controlar la capacidad de la red: probaremos que lo mayor es la componente a¡, lo menor es el número de puntos fijos pertenecientes a la línea de energía R¡. Una vez que la capacidad de la red ha sido controlada mediante este parámetro, introducimos otro parámetro, definido como la desviación del vector de pesos relativos, este parámetro sirve para disminuir ostensiblemente el número de parásitos. A lo largo de todo el trabajo, hemos ido desarrollando un ejemplo, el cual nos ha servido para ir corroborando los resultados teóricos, los algoritmos están escritos en un pseudocódigo, aunque a su vez han sido implamentados utilizando el paquete Mathematica 2.2., mostrándolos en un volumen suplementario al texto.---ABSTRACT---The aim of the present research is intended to explore new specifícation techniques of Neural Networks based on Graphs to be used in the optimization and simplification of Network Architectures and Computational Complexhy. We have focused our attention in a, well known, class of Neural Networks: the Recursive Neural Networks, also known as Hopfield's Neural Networks. The general problem of constructing the synaptic matrix associated with a Recursive Neural Network imposing some vectors as fixed points is fer for completery solved, the number of prototype vectors (learning patterns) which can be stored by Hebb's law is rather limited and the memory will thus quickly reach saturation if new prototypes are continuously acquired in the course of time. Hebb's law needs thus to be revised in order to allow new prototypes to be stored at the expense of the older ones. Some approaches related with this problem has been developed. We have developed a new approach of implementing a Recursive Neural Network in order to sob/e these kind of problems, the synaptic matrix is obtained superposing the components of the prototype vectors over the vértices of a Graph which may be interpreted as a coloring of the Graph. When training is finished the adjacency matrix of the Resulting Graph or matrix of weights presents certain properties for which it may be called a tetrahedral matrix The energy associated to any possible state of the net is represented as a point (a,b) in R2. Every one of the energy points associated with state-vectors having the same Hamming distance to the zero vector are located over the same energy Une in R2. The state-vector space may be then classified in n classes according to the n different possible distances firom any of the state-vectors to the zero vector The (n x n) matrix of weights may also be reduced to a n-vector of weights, in this way the computational time and the memory space required for obtaining the weights is optimized and simplified. In the recall stage, a parameter vectora is introduced, this parameter is used for controlling the capacity of the net: it may be proved that the bigger is the r, component of J, the lower is the number of fixed points located in the r¡ energy line. Once the capacity of the net has been controlled by the ex parameter, we introduced other parameter, obtained as the relative weight vector deviation parameter, in order to reduce the number of spurious states. All along the present text, we have also developed an example, which serves as a prove for the theoretical results, the algorithms are shown in a pseudocode language in the text, these algorithm so as the graphics have been developed also using the Mathematica 2.2. mathematical package which are shown in a supplementary volume of the text.

Extended phone log-likelihood ratio features and acoustic-based I-vectors for language recognition

This paper presents new techniques with relevant improvements added to the primary system presented by our group to the Albayzin 2012 LRE competition, where the use of any additional corpora for training or optimizing the models was forbidden. In this work, we present the incorporation of an additional phonotactic subsystem based on the use of phone log-likelihood ratio features (PLLR) extracted from different phonotactic recognizers that contributes to improve the accuracy of the system in a 21.4% in terms of Cavg (we also present results for the official metric during the evaluation, Fact). We will present how using these features at the phone state level provides significant improvements, when used together with dimensionality reduction techniques, especially PCA. We have also experimented with applying alternative SDC-like configurations on these PLLR features with additional improvements. Also, we will describe some modifications to the MFCC-based acoustic i-vector system which have also contributed to additional improvements. The final fused system outperformed the baseline in 27.4% in Cavg.

State space models for multi-setup operational modal analysis

Operational Modal Analysis consists on estimate the modal parameters of a structure (natural frequencies, damping ratios and modal vectors) from output-only vibration measurements. The modal vectors can be only estimated where a sensor is placed, so when the number of available sensors is lower than the number of tested points, it is usual to perform several tests changing the position of the sensors from one test to the following (multiple setups of sensors): some sensors stay at the same position from setup to setup, and the other sensors change the position until all the tested points are covered. The permanent sensors are then used to merge the mode shape estimated at each setup (or partial modal vectors) into global modal vectors. Traditionally, the partial modal vectors are estimated independently setup by setup, and the global modal vectors are obtained in a postprocess phase. In this work we present two state space models that can be used to process all the recorded setups at the same time, and we also present how these models can be estimated using the maximum likelihood method. The result is that the global mode shape of each mode is obtained automatically, and subsequently, a single value for the natural frequency and damping ratio of the mode is computed. Finally, both models are compared using real measured data.