AI Hot Machine Learning & Data Science Nature Language Tech Research

Google ‘BigBird’ Achieves SOTA Performance on Long-Context NLP Tasks

BigBird is shown to dramatically improve performance across long-context NLP tasks, producing SOTA results in question answering and summarization.

The phenomenal success of Google’s BERT and other natural language processing (NLP) models based on transformers isn’t accidental. Behind all the SOTA performances lies transformers’ innovative self-attention mechanism, which enables networks to capture contextual information from an entire text sequence. However, the memory and computational requirements of self-attention grow quadratically with sequence length, making it very expensive to use transformer-based models for processing long sequences.

To alleviate the quadratic dependency of transformers, a team of researchers from Google Research recently proposed a new sparse attention mechanism dubbed BigBird. In their paper Big Bird: Transformers for Longer Sequences, the team demonstrates that despite being a sparse attention mechanism, BigBird preserves all known theoretical properties of quadratic full attention models. In experiments, BigBird is shown to dramatically improve performance across long-context NLP tasks, producing SOTA results in question answering and summarization.

image.png

The researchers designed BigBird to satisfy all known theoretical properties of full transformers, building three main components into the model:

  • A set of g global tokens that attend to all parts of a sequence.
  • For each query qi , a set of r random keys that each query will attend to.
  • A block of local neighbours w so that each node attends on their local structure

These innovations enable BigBird to handle sequences up to eight times longer than what was previously possible using standard hardware.

image.png

Additionally, inspired by the capability of BigBird to handle long contexts, the team introduced a novel application of attention-based models for extracting contextual representations of genomics sequences like DNA. In experiments, BigBird proved to be beneficial in processing the longer input sequences and also delivered improved performance on downstream tasks such as promoter-region and chromatin profile prediction.

image.png
image.png
image.png

The paper Big Bird: Transformers for Longer Sequences is on arXiv.


Reporter: Fangyu Cai | Editor: Michael Sarazen


Image for post

Synced Report | A Survey of China’s Artificial Intelligence Solutions in Response to the COVID-19 Pandemic — 87 Case Studies from 700+ AI Vendors

This report offers a look at how the Chinese government and business owners have leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon Kindle.

Click here to find more reports from us.


We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.

%d bloggers like this: