Visualizing internal mechanisms
We have seen the inner workings of the transformer, how it can be trained, and the main types of models. The beauty of attention is that we can visualize these relationships, and in this section, we will see how to do that. We can then visualize the relationships within the BERT attention head. As mentioned, in each layer, there are several attention heads and each of them learns a different representation of the input data. The color intensity indicates a greater weight in the attention weights (darker colors indicate weights that are close to 1).
We can do this using the BERTviz package:
head_view(attention, tokens, sentence_b_start)
Important note
The visualization is interactive. The code is in the repository. Try running it using different phrases and exploring different relationships between different words in the phrases. The visualization allows you to explore the different layers in the model by taking advantage of the drop-down model...
