GradientExplainer

New 'interior interpretability' method probes Transformer models

Researchers have introduced a new concept called "interior interpretability" to better understand the internal workings of Transformer models. This approach uses attention rollout, viewing it as an operator that mediate…