Due to recent breakthroughs, current learning systems are able to overperform human’s
capabilities in a lot of fields, such as image processing, natural language processing, and
voice processing. However, this performance requires a lot of training data and huge
computing resources. The ultimate goal of my work is to create a learning system that
can efficiently use training data. Efficiency is defined by the number of training data. The
problem can be studied from a variety of approaches, the author discusses two solutions:
the one-shot learning and the active learning direction are presented in this thesis.
In the first part of this work, I examine the possibilities of creating an architecture that uses
a very few data points and this system is able to classify with minimal accuracy decrease.
I study the one-shot learning task and optimize the learning system. In this task, the
model can use exactly one sample per class to learn the task. With this constraint, the
problem was shifted towards extracting the useful features of the data points. During this
task, these features are also learned from another dataset with similar characteristics.
The second half of this work attempts to optimize the problem of collecting data in the
way of controlling the collecting process to get more valuable data. A data point can be
considered a more valuable data if it determines the decision more than another sample.
This issue is part of the active learning topic. During the thesis, I study a situation in
which the data generator is available and the properties of the desired data point can be
adjusted. This task is illustrated by an example of the chapter in which speaker recognition
is the task and data generators are the persons to be recognized. The spoken word itself
can be controlled by determining the word itself and the speaker’s identity