Enhancing Robustness and Interpretability in Transformer Networks through Fuzzification

Transformer networks have shown good performance in multifaceted tasks such as natural language processing, but face challenges in terms of robustness and interpretability. This study proposes a Transformer Fuzzy Neural Network (TFNN) that integrates fuzzification into the traditional transformer architecture. By incorporating fuzzy logic, TFNN enhances model stability and interpretability, enabling better handling of ambiguous data. Experiments conducted on the WMT 2014 English-German dataset demonstrate that TFNN outperforms the classical transformer, achieving a BLEU score of 23.2. This research highlights the potential of fuzzification to improve AI model robustness and interpretability, paving the way for broader applications in critical areas.