18696052. Learning the Joint Distribution of Two Sequences Using Little or No Paired Data (Google LLC)
Learning the Joint Distribution of Two Sequences Using Little or No Paired Data
Organization Name
Inventor(s)
Soroosh Mariooryad of Redwood City CA US
Sean Matthew Shannon of San Francisco CA US
Thomas Edward Bagby of Monte Rio CA US
David Teh-Hwa Kao of Philadelphia PA US
Daisy Antonia Stanton of San Francisco CA US
Eric Dean Battenberg of Walnut Creek CA US
Russell John Wyatt Skerry-ryan of Mountain View CA US
Learning the Joint Distribution of Two Sequences Using Little or No Paired Data
This abstract first appeared for US patent application 18696052 titled 'Learning the Joint Distribution of Two Sequences Using Little or No Paired Data
Original Abstract Submitted
Provided is a noisy channel generative model of two sequences, for example text and speech, which enables uncovering the associations between the two modalities when limited paired data is available. To address the intractability of the exact model under a realistic data set-up, example aspects of the present disclosure include a variational inference approximation. To train this variational model with categorical data, a KL encoder loss approach is proposed which has connections to the wake-sleep algorithm.
- Google LLC
- Soroosh Mariooryad of Redwood City CA US
- Sean Matthew Shannon of San Francisco CA US
- Thomas Edward Bagby of Monte Rio CA US
- Siyuan Ma of San Jose CA US
- David Teh-Hwa Kao of Philadelphia PA US
- Daisy Antonia Stanton of San Francisco CA US
- Eric Dean Battenberg of Walnut Creek CA US
- Russell John Wyatt Skerry-ryan of Mountain View CA US
- G06N3/088
- G06N3/0455
- CPC G06N3/088
(Ad) Transform your business with AI in minutes, not months
Trusted by 1,000+ companies worldwide
