Postgrado - Ciencias de la Computación

Permanent URI for this community

https://hdl.handle.net/20.500.12590/17962

Browse

Browsing Postgrado - Ciencias de la Computación by browse.metadata.advisor "Mora Colque, Rensso Victor Hugo"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Estimación de la incertidumbre en redes neuronales profundas
    (Universidad Católica San Pablo, 2020) Jaita Aguilar, Jose Hugo; Mora Colque, Rensso Victor Hugo
    Hoy en día, los modelos de aprendizaje profundo representan el estado del arte en muchas tareas, lo que ha motivado a utilizarse en distintas aplicaciones reales. Varias de ellas se encuentran en campos como: la medicina, seguridad, finanzas, etc. en donde una predicción errónea puede llegar a ser fatal. Por lo tanto, necesitamos que los modelos no solo den una predicción, sino que además un grado de certeza acerca de ella, es decir, la incertidumbre en la predicción. En esta tesis se estudian dos tipos de incertidumbre: la epistémica (la cual captura la falta de certeza del modelo) y la aleatoria (generada por el ruido en los datos). La estimación de la incertidumbre epistémica es un desafío, siendo el enfoque bayesiano el más utilizado para abordarla (debido a las herramientas que nos ofrece), pero este viene con un costo computacional prohibitivo, evidenciado aún más en modelos de aprendizaje profundo. Nosotros proponemos el método SVGD-A, utilizando como base al método SVGD, para realizar la inferencia bayesiana. Nuestra propuesta se enfoca en acelerar el proceso de convergencia de SVGD, permitiendo el escalamiento a modelos profundos. En cuanto a la incertidumbre aleatoria, nosotros proponemos un método basado en la extracción de características de bajo nivel en modelos ya entrenados, para luego aplicarles (a las características) una reducción de dimensionalidad con t-SNE, volviendo as ́ı el problema en una tarea de cauterización. Además, proponemos un segundo método el cual utiliza el mismo esquema descrito anteriormente, pero con la novedad que se le agrega un Autoencoder Variacional. Y por último, aportamos con un novedoso enfoque para realizar inferencia usando t-SNE. Finalmente, mostramos la eficiencia de nuestros métodos en la tarea de detectar muestras out-of-distribution en distintas bases de datos, logrando resultados muy superiores al estado del arte.
  • Thumbnail Image
    Item
    Priority sampling and visual attention for self-driving car
    (Universidad Católica San pablo, 2023) Flores Benites, Victor; Mora Colque, Rensso Victor Hugo
    End-to-end methods facilitate the development of self-driving models by employing a single network that learns the human driving style from examples. However, these models face problems such as distributional shift, causal confusion, and high variance. To address these problems we propose two techniques. First, we propose the priority sampling algorithm, which biases a training sampling towards unknown observations for the model. Priority sampling employs a trade-off strategy that incentivizes the training algorithm to explore the whole dataset. Our results show a reduction of the error in the control signals in all the models studied. Moreover, we show evidence that our algorithm limits overtraining on noisy training samples. As a second approach, we propose a model based on the theory of visual attention (Bundesen, 1990) by which selecting relevant visual information to build an optimal environment representation. Our model employs two visual information selection mechanisms: spatial and feature-based attention. Spatial attention selects regions with visual encoding similar to contextual encoding, while feature-based attention selects features disentangled with useful information for routine driving. Furthermore, we encourage the model to recognize new sources of visual information by adding a bottom-up input. Results in the CoRL-2017 dataset (Dosovitskiy et al., 2017) show that our spatial attention mechanism recognizes regions relevant to the driving task. Our model builds disentangled features with low cosine similarity, but with high representation similarity. Finally, we report performance improvements over traditional end-to-end models.
  • Thumbnail Image
    Item
    Surveillance video summarization based on trajectory rarity measure
    (Universidad Católica San Pablo, 2019) Quispe Torres, Gerar Francis; Mora Colque, Rensso Victor Hugo
    The dynamic video summarization of surveillance videos has several critical applications, mainly due to the wide availability of digital cameras in environments such as airports, train and bus stations, shopping centers, stadiums, buildings, schools, hospitals, roads, among others. This study presents an approach for the generation of dynamic summary on surveillance video domain based on human trajectories. It has an emphasis on trajectory descriptors in conjunction with the unsupervised clustering method. Our approach contribute to existing literature concerning the combination of methods and objectives. We hypothesize that the clustering of trajectories permits to identify rare trajectories base on their morphology. The clustering as an output provides numerous subsets of trajectories or clusters and the number of elements of a specific cluster is used to determine their rarity. Those subsets with few components are rare while the others that have a high number of elements are considered ordinary; therefore, the implications of our study show that is possible to use unsupervised clustering for automatic detection of rare trajectories based on their morphology and with this information segment videos. We experimented with different sets of trajectories segmenting the rare videos from our ground truth.