Alexa Blogs

Alexa Blogs

Want the latest?

alexa topics

Recent Posts

Archive

Showing posts tagged with Alexa Science

March 21, 2019

Rahul Gupta

To produce synthetic training data for a machine-learning application where real data is scarce, Alexa scientist Rahul Gupta uses generative adversarial learning, which pits two neural nets against each other -- one trying to generate convincing fakes, the other trying to discern fake from real.

[Read More]

March 20, 2019

Minhua Wu

Using one neural network to label speech data, adding synthetic environmental noise to that data, and then using it to train a second neural network improves speech recognition, particularly under noisy conditions.

[Read More]

March 11, 2019

Ming Sun

Imbalances in the data used to train machine learning systems can cause biases; a new method for correcting imbalances increases the accuracy of a sound detection system by 22% over systems that used the previous method.

[Read More]

March 05, 2019

Behnam Hedayatnia

As the Alexa Prize Socialbot Challenge 3 gears up, applied scientist Behnam Hedayatnia reviews a few of the innovations from the university teams who participated in the previous challenge.

[Read More]

February 27, 2019

Anu Venkatesh

Tools include a contextual topic model, a sensitive-content detector, a conversation evaluator, and an integrated development environment. 

[Read More]

January 31, 2019

Mike Rodehorst

"Acoustic fingerprinting" tells Alexa when specific instances of her name are safe to ignore.

[Read More]

January 30, 2019

Alessandro Moschitti

A novel transfer-learning technique enables the addition of new classification categories to an existing machine learning model without access to the model's original training data.

[Read More]

January 24, 2019

Alessandro Moschitti

New method for comparing data structures enables natural-language-understanding system to learn in four hours what used to take more than seven days.

[Read More]

January 22, 2019

Anuj Goyal

A new approach leverages hundreds of millions of unannotated Alexa requests to improve the quality of "transfer learning", or adapting an existing neural network, trained on abundant data, to a new task for which training data is scarce.

[Read More]

January 15, 2019

Anish Acharya

By compressing the huge lookup tables that list "embeddings", or vector representations of individual words, a new system can shrink neural-network models by up to 90%, with minimal effect on accuracy.

[Read More]

December 21, 2018

Rasool Fakoor

Personal memory retrieval systems are usually trained to do one thing then judged on their ability to do something else. Alexa scientists show how reinforcement learning lets them use the same criterion (F1 score) to both train and evaluate such systems.

[Read More]

December 18, 2018

Sanchit Agarwal

A new parser that learns which words of an utterance belong together enables Alexa to handle multistep requests.

[Read More]

December 17, 2018

Young-Bum Kim

Data representation schemes — such as "embeddings" — are a hot topic in machine learning. At this year's IEEE Spoken Language Technologies conference, Alexa scientists present a new representation scheme that cuts skill selection error rate by 40%.

[Read More]

December 13, 2018

Ankur Gandhe

New Alexa capabilities are often bootstrapped using "grammars", formal rules that can generate artificial training examples for machine learning systems. A new method for constructing statistical language models directly from grammars can improve speech recognition on new capabilities by up to 15%.

[Read More]

December 11, 2018

Chengwei Su

A new technique that lets different natural-language-understanding modules independently calibrate their judgments about whether they should handle a given utterance helps improvements reach Alexa customers more efficiently.

[Read More]

Want the latest?

alexa topics

Recent Posts

Archive