Timezone: »

Influence Patterns for Explaining Information Flow in BERT
Kaiji Lu · Zifan Wang · Piotr Mardziel · Anupam Datta

Fri Dec 10 08:30 AM -- 10:00 AM (PST) @

While attention is all you need may be proving true, we do not know why: attention-based transformer models such as BERT are superior but how information flows from input tokens to output predictions are unclear. We introduce influence patterns, abstractions of sets of paths through a transformer model. Patterns quantify and localize the flow of information to paths passing through a sequence of model nodes. Experimentally, we find that significant portion of information flow in BERT goes through skip connections instead of attention heads. We further show that consistency of patterns across instances is an indicator of BERT’s performance. Finally, we demonstrate that patterns account for far more model performance than previous attention-based and layer-based methods.

Author Information

Kaiji Lu (Carnegie Mellon University)
Zifan Wang (Carnegie Mellon University)
Piotr Mardziel (Carnegie Mellon University)
Anupam Datta (Carnegie Mellon University)

More from the Same Authors