The field of artificial intelligence has been throughout its history repeatedly inspired by human cognition. In this master’s thesis, we take a high-level view of meta-cognition and implement a system with similar characteristics. Our motivation for this is of two kinds: the first stems from the desire to contribute to machine learning methods, more specifically semi-supervised classification, and the second from the ability to compare human learning and artificial systems. According to the literature, our system is divided into object and meta parts, with the former solving the classification problem, and the latter selecting learning examples based on its knowledge by determining the appropriate confidence thresholds. Similarly to humans, we learn learning strategies in our system through the accumulation of knowledge about solving a particular problem, for which we use reinforcement learning. While designing our system, one of the important guidelines is generality, which is why we perform experiments by varying the architectures of classifiers (neural networks) and datasets. We train the system both from the beginning and by transferring knowledge from one problem to another. We obtain mixed results that depend largely on the effectiveness of our approach to semi-supervised learning. When comparing our method with “naïve” approaches, we get at most 1% worse, but often better results than using confidence thresholds found by random search. By transferring learning strategies from one problem to another, we reduce the time required to solve the sample selection problem by 80% and obtain comparable results as when learning from the beginning. Our work shows that it is possible to learn a curriculum within a given framework, that it accelerates learning and that the selection of learning samples using meta-knowledge is one of the effective approaches for successful classifier training. The listed properties are also key similarities of the implemented system when compared to human learning.
|
