Influence Patterns for Explaining Information Flow in BERT
Dec 6, 2021
Speakers
Zifan Wang
Speaker · 0 followers
Piotr Mardziel
Speaker · 0 followers
Anupam Datta
Speaker · 0 followers
About
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.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 throug…
Organizer
NeurIPS 2021
Account · 1.9k followers
About NeurIPS 2021
Neural Information Processing Systems (NeurIPS) is a multi-track machine learning and computational neuroscience conference that includes invited talks, demonstrations, symposia and oral and poster presentations of refereed papers. Following the conference, there are workshops which provide a less formal setting.
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
Noether Networks: Meta-Learning Useful Conserved Quantities
Watch later
Ferran Alet, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Differentiable Equilibrium Computation with Decision Diagrams for Stackelberg Models of Combinatorial Congestion Games
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Distributed Machine Learning with Sparse Heterogeneous Data
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Panel Discussion 3
Watch later
Taylor Webb, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Competition Track Day 1
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Scheduling jobs with stochastic holding costs
Watch later
Dabeen Lee, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%