Publication: Herramienta de visualización del aprendizaje en redes de neuronas
Loading...
Identifiers
Publication date
2013-10-31
Defense date
2013-10-31
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
En este proyecto se ha desarrollado una aplicación que permite visualizar el proceso de aprendizaje de una red de neuronas. Visualizando la evolución temporal de la frontera de separación entre los datos, el proceso de minimización del error... Se ha desarrollado teniendo como “target” el uso docente; Buscando facilitar para la enseñanza universitaria una herramienta en la que se pueda mostrar gráficamente algunos de los distintos aspectos de las redes de neuronas, de forma simplificada. Las gráficas de la aplicación se presentan en una representación 2D por ser la forma más simple y práctica para representar gráficamente una red de neuronas; Además de forma consensuada con los tutores se añadió la restricción de trabajar con un único perceptrón que tendrá una capa de entrada de dos neuronas, una única capa oculta de dos neuronas, y una capa de salida de una neurona. La aplicación una vez configurada, se inicia ejecutando el entrenamiento de la red de neuronas, buscando optimizar el valor de los pesos, y en el que se incluye el algoritmo de propagación hacia atrás, mostrando los resultados que va produciendo cada ciclo a tiempo real en la interfaz principal; Dichos resultados se visualizan en dos gráficas que mostrarán distintas representaciones de la ejecución; Una de ellas corresponderá a la representación utilizando los ejes naturales, y la segunda de ellas tendrá como ejes los valores de salida de las dos neuronas de la capa oculta; Mostrando en ambos casos en cada punto el valor de la salida final de la red. La aplicación además permitirá la interacción con el usuario durante la ejecución del entrenamiento; Además otras funcionalidades que servirán para profundizar más en detalle de la evolución producida con el algoritmo de propagación hacia atrás. Estas funcionalidades son:
- La posibilidad de parar la ejecución y seleccionar un ciclo de ejecución concreto lo cual permitirá observar de una forma distinta y más detenidamente la evolución que han ido sufriendo las gráficas.
- Mostrar un nuevo gráfico que presente la estructura del perceptrón gráficamente mostrando las neuronas, las conexiones, y los valores de los pesos de cada una de las conexiones en el ciclo concreto en el que se inicie la funcionalidad.
- Mostrar la evolución del error durante toda la ejecución, permitiendo observar mediante una “gráfica de línea” la evolución del error presentadas según van avanzando los ciclos.
- Gráficas que muestran las variaciones del error en función de pequeñas variaciones en el peso de las conexiones de neuronas concretas.
This project has developed an application that displays the learning process of a neural network Visualizing the evolution of the boundary of separation between the data, the error minimization process ... It´s also important to say that is developed for educational use; Looking for simplify the teaching in the university using a tool which can show some of the different aspects of neural networks, in a simplified way. The application its presented in a 2D representation because it was the simplest, and most practical way to “graph” a neural network; Also in consensus with tutors I also added the restriction of working with a single perceptron, which will have two neurons in the input layer, a single hidden layer with also two neurons and finally an output layer with one neuron. The application once configured, it starts running the training of the neural network, seeking to optimize the value of the weights, and which includes the algorithm back propagation, showing the results of each cycle in real time on the main interface; These results are displayed in two charts that show different representations of the execution; The first one correspond to the representation using the natural axes, and the second one is based on using as axes the output values of the two neurons in the hidden layer; Showing in both cases at each point the output value of the neural network. The application allows user interaction during the execution of the training. Also offers other features that are used to deepen more in detail on the evolution produced by the back propagation algorithm. These features are: - The possibility to stop the execution, and select a previous execution cycle which will allow a different and closer way to look at the evolution happened in the graphs. - Displaying a new graph that present the perceptron´s structure graphically, displaying the neurons, and links, and the values of the weights of each one of the connections in the particular cycle in which we are situated. - The evolution of the error during the entire execution, allowing us to observe error evolution, using a "line graph", while we are moving forward cycles. - Graphs showing variations of the global error depending on small variations in the value of the weight in connections of a concrete neuron.
This project has developed an application that displays the learning process of a neural network Visualizing the evolution of the boundary of separation between the data, the error minimization process ... It´s also important to say that is developed for educational use; Looking for simplify the teaching in the university using a tool which can show some of the different aspects of neural networks, in a simplified way. The application its presented in a 2D representation because it was the simplest, and most practical way to “graph” a neural network; Also in consensus with tutors I also added the restriction of working with a single perceptron, which will have two neurons in the input layer, a single hidden layer with also two neurons and finally an output layer with one neuron. The application once configured, it starts running the training of the neural network, seeking to optimize the value of the weights, and which includes the algorithm back propagation, showing the results of each cycle in real time on the main interface; These results are displayed in two charts that show different representations of the execution; The first one correspond to the representation using the natural axes, and the second one is based on using as axes the output values of the two neurons in the hidden layer; Showing in both cases at each point the output value of the neural network. The application allows user interaction during the execution of the training. Also offers other features that are used to deepen more in detail on the evolution produced by the back propagation algorithm. These features are: - The possibility to stop the execution, and select a previous execution cycle which will allow a different and closer way to look at the evolution happened in the graphs. - Displaying a new graph that present the perceptron´s structure graphically, displaying the neurons, and links, and the values of the weights of each one of the connections in the particular cycle in which we are situated. - The evolution of the error during the entire execution, allowing us to observe error evolution, using a "line graph", while we are moving forward cycles. - Graphs showing variations of the global error depending on small variations in the value of the weight in connections of a concrete neuron.
Description
Keywords
Redes neuronales, Inteligencia artificial, Aprendizaje, Desarrollo de software, Interfaces gráficas de usuario