Summary
In this chapter, we discussed one of the most important problems in front of AI: XAI. Explainability becomes a serious issue as language models grow with a strong trend. We addressed two topics in terms of Transformers only. First, we examined the self-attention mechanism in the architecture. We tried to understand the internal process of these mechanisms with various visualization tools. Secondly, we interpreted the decision-making process of Transformer architectures. We worked with two model-agnostic approaches: LIME and SHAP. With these two techniques, we monitor how models give importance to parts of inputs (words) in a simple text classification process.
In the next chapter, we will focus on a special kind of transformer namely the efficient transformer.
