Transformer Attention Visualisation — Explainability & Analysis

I will implement attention visualisation and explainability analysis for your transformer model — covering attention head pattern analysis, Integrated Gradients attribution, BERTViz attention flow visualisation, probing classifier analysis for understanding what information is encoded in different layers, and a report explaining what your transformer model has learned in language accessible to non-ML stakeholders. Transformer attention is frequently cited as providing interpretability, but raw attention weights are not reliable importance scores — the implementation here uses gradient-based attribution methods that are theoretically grounded, alongside attention visualisation as a complementary analysis tool.

The analysis covers attention head pattern clustering, Integrated Gradients importance attribution, BERTViz visualisation for text transformers, probing analysis for specific linguistic properties, and an interpretability report with findings in plain language.