• Home
  • About Us
  • Contact Us
  • Privacy Policy
  • Special Offers
Business Intelligence Info
  • Business Intelligence
    • BI News and Info
    • Big Data
    • Mobile and Cloud
    • Self-Service BI
  • CRM
    • CRM News and Info
    • InfusionSoft
    • Microsoft Dynamics CRM
    • NetSuite
    • OnContact
    • Salesforce
    • Workbooks
  • Data Mining
    • Pentaho
    • Sisense
    • Tableau
    • TIBCO Spotfire
  • Data Warehousing
    • DWH News and Info
    • IBM DB2
    • Microsoft SQL Server
    • Oracle
    • Teradata
  • Predictive Analytics
    • FICO
    • KNIME
    • Mathematica
    • Matlab
    • Minitab
    • RapidMiner
    • Revolution
    • SAP
    • SAS/SPSS
  • Humor

Google’s Evolved Transformer achieves state-of-the-art performance in translation tasks

June 16, 2019   Big Data

The Transformer, a type of AI architecture introduced in a 2017 paper (“Attention Is All You Need“) coauthored by scientists at Google, excels at writing prose and product reviews, synthesizing voices, and crafting harmonies in the style of classical composers. But a team of Google researchers believed it could be taken a step further with AutoML, a technique in which a “controller” system identifies a “child” architecture that can then be tailored to a particular task. Remarkably, the result of their work — which they describe in a newly published paper and accompanying blog post — achieves both state-of-the-art translation results and improved performance on language modeling compared with the original Transformer.

They’ve released the new model — Evolved Transformer — as part of Tensor2Tensor, a library of open source AI models and data sets.

Traditionally, AutoML approaches begin with a pool of random models that the controller trains and evaluates for quality. The process is repeated thousands of times, and each time results in new vetted machine learning architectures from which the controller learns. Eventually, the controller begins to assign high probability to model components that achieve better accuracy on validation data sets and low probability to poorly scoring areas.

Discovering the Evolved Transformer with AutoML necessitated the development of two new techniques, the researchers say, because the task used to evaluate the performance of each architecture (WMT’14 English-German translation) was computationally expensive. The first — warm starting — seeded the initial model population with the Transformer architecture instead of random models, which helped ground the search. Meanwhile, the second — Progressive Dynamic Hurdles (PDH) — augmented the search to allocate more resources to the strongest candidates, enabling the controller to terminate the evaluation of “flagrantly bad” models early and award promising architectures more resources.

 Google’s Evolved Transformer achieves state of the art performance in translation tasks

Above: The Evolved Transformer architecture.

Image Credit: Google AI

So what’s so special about the Evolved Transformer? As with all deep neural networks, the Evolved Transformer contains neurons (functions) that transmit “signals” from input data and slowly adjust the synaptic strength — weights — of each connection, which is how the model extracts features and learns to make predictions. Furthermore, the Evolved Transformer has attention, such that every output element is connected to every input element and the weightings between them are calculated dynamically.

Like most sequence-to-sequence models, the Evolved Transformer contains an encoder that encodes input data (sentences in translation tasks) into embeddings (mathematical representations) and a decoder that uses those embeddings to construct outputs (translations).

But the team notes that it contains something rather unconventional, as well: convolutional layers at the bottom of both the encoder and decoder modules in branching pattern, such that inputs run through two separate convolutional layers before being added together. Whereas the original Transformer relied solely on attention, then, the Evolved Transformer is a hybrid that leverages the strengths of both self-attention and wide convolution.

 Google’s Evolved Transformer achieves state of the art performance in translation tasks

Above: The Evolved Transformer’s performance compared with the Transformer.

Image Credit: Google AI

In tests, the team compared the Evolved Transformer with the original Transformer on the English-German translation task used during the model search, and found that the former achieved better performance on both BLEU (an algorithm for evaluating the quality of machine-translated text) and perplexity (a measurement of how well probability distribution predicts a sample) at all sizes. At larger sizes, the Evolved Transformer reached state-of-the-art performance with a BLEU score of 29.8, and on experiments involving translation with different language pairs and language modeling, they observed a performance improvement of nearly two perplexity.

Let’s block ads! (Why?)

Big Data – VentureBeat

achieves, Evolved, Google's, Performance, stateoftheart, Tasks, Transformer, translation
  • Recent Posts

    • Accelerate Your Data Strategies and Investments to Stay Competitive in the Banking Sector
    • SQL Server Security – Fixed server and database roles
    • Teradata Named a Leader in Cloud Data Warehouse Evaluation by Independent Research Firm
    • Derivative of a norm
    • TODAY’S OPEN THREAD
  • Categories

  • Archives

    • April 2021
    • March 2021
    • February 2021
    • January 2021
    • December 2020
    • November 2020
    • October 2020
    • September 2020
    • August 2020
    • July 2020
    • June 2020
    • May 2020
    • April 2020
    • March 2020
    • February 2020
    • January 2020
    • December 2019
    • November 2019
    • October 2019
    • September 2019
    • August 2019
    • July 2019
    • June 2019
    • May 2019
    • April 2019
    • March 2019
    • February 2019
    • January 2019
    • December 2018
    • November 2018
    • October 2018
    • September 2018
    • August 2018
    • July 2018
    • June 2018
    • May 2018
    • April 2018
    • March 2018
    • February 2018
    • January 2018
    • December 2017
    • November 2017
    • October 2017
    • September 2017
    • August 2017
    • July 2017
    • June 2017
    • May 2017
    • April 2017
    • March 2017
    • February 2017
    • January 2017
    • December 2016
    • November 2016
    • October 2016
    • September 2016
    • August 2016
    • July 2016
    • June 2016
    • May 2016
    • April 2016
    • March 2016
    • February 2016
    • January 2016
    • December 2015
    • November 2015
    • October 2015
    • September 2015
    • August 2015
    • July 2015
    • June 2015
    • May 2015
    • April 2015
    • March 2015
    • February 2015
    • January 2015
    • December 2014
    • November 2014
© 2021 Business Intelligence Info
Power BI Training | G Com Solutions Limited