Loosely Symmetric Neural Network Implementation

Insufficient or imbalanced data are a common issue, e.g., in healthcare where the data is limited to protect personal data and privacy. This imbalance results in classifiers' poor performance as they usually prioritize the majority class during classification. The artificial neural network requires a large amount of well-prepared balanced data. Human beings, on the other hand, can extract new information even from a few. This thesis deals with the implementation of cognitive biases into neural networks, which should help the model imitate human learning. The loosely symmetric model designed by Taniguchi et al. had been tested before by authors in the language R and achieved the best results from all tested models. During this thesis, the implementation was created in Python as there is no other available. The thesis contains a theoretical introduction to machine learning, the fit algorithm of the neural network, the loosely symmetric model description, including created variants, and the evaluation. During the evaluation of the spam classification, the new LSNN implementation managed to keep up with the model LSNB and even achieved better performance than the eLSNB model, which performed best on the same dataset in previous research. However, the method of "enhancement" is not clearly explained in the original description. Therefore, it is unknown whether this model is the correct implementation of the original design of the LSNN model. This implementation of the LSNN model proved its potential. However, the fit process could be further extended with the method of early stopping.