Învățare adâncă pentru aplicații cu date insuficiente / Deep Learning for scarce data applications

This thesis addresses the problem of scarce data applications. We propose two different solutions for data processing. (1) We extract fundamental information from the provided datasets. We present 3 lifelog moment retrieval system that focus on removing images that do not match the query proposed by the user and finding measures of similarity between images and textual description based on meta-information and image content. We propose a similar approach in the case of an abandoned luggage detection system, where we use the descriptors of an object detector to solve three different tasks: abandoned luggage detection, detection of the person who left the luggage and re-identification of that person in a CCTV stream. (2) We augment datasets using unsupervised generative algorithms. Thus, we propose a logo generation system starting from an initial image, we pass it through a gradient backpropagation algorithm to extract the latent vector used for its generation and we obtain an approximate reconstruction of the initial logo with which we augment the dataset. We validate the augmented dataset with the help of a logo detection system. In the last part of the thesis we focus on the generation of financial time series. We propose 14 different major architectures (4 MLP, 6 FCGAN, 2 VAE, 2 GMMN) that we use to generate fixed-length financial time series. Then, we propose a mechanism for combining these sequences to reach arbitrary lengths. We propose and study a wide range of qualitative and quantitative metrics inspired by information theory to determine the “realism” of synthetic samples and validate our approach through a regime prediction mechanism that will predict a time series’ behaviour in the near future.