Domain adaptation by multi-noising stacked marginalized denoising encoders

Patent metadata

Field	Value
Publication number	US-9916542-B2
Application number	US-201615013273-A
Country	US
Kind code	B2
Filing date	Feb 2, 2016
Priority date	Feb 2, 2016
Publication date	Mar 13, 2018
Grant date	Mar 13, 2018

Abstract

A machine learning method operates on training instances from a plurality of domains including one or more source domains and a target domain. Each training instance is represented by values for a set of features. Domain adaptation is performed using stacked marginalized denoising autoencoding (mSDA) operating on the training instances to generate a stack of domain adaptation transform layers. Each iteration of the domain adaptation includes corrupting the training instances in accord with feature corruption probabilities that are non-uniform over at least one of the set of features and the domains. A classifier is learned on the training instances transformed using the stack of domain adaptation transform layers. Thereafter, a label prediction is generated for an input instance from the target domain represented by values for the set of features by applying the classifier to the input instance transformed using the stack of domain adaptation transform domains.

First claim

The invention claimed is:

1. A device comprising:
a computer programmed to perform a machine learning method operating on training instances from a plurality of domains including one or more source domains and a target domain, with each training instance represented by values for a set of features wherein each domain corresponds to a traffic light surveillance camera and the training instances of each domain are images acquired by the corresponding surveillance camera of vehicles and the traffic light monitored by the corresponding surveillance camera, the set of labels indicate whether the imaged vehicle is running a red light, and the classifier outputs label predictions as to whether an imaged vehicle is running the traffic light, the machine learning method including the operations of:
performing domain adaptation using stacked marginalized denoising autoencoding (mSDA) operating on the training instances to generate a stack of domain adaptation transform layers wherein each iteration of the domain adaptation includes corrupting the training instances in accord with feature corruption probability vectors for the features that are non-uniform over the domains including:
corrupting the training instances from each source domain in accord with a feature corruption probability specific to that source domain; and
corrupting the training instances from the target domain in accord with a feature corruption probability specific to the target domain;

learning a classifier on the training instances transformed using the stack of domain adaptation transform layers; and
generating a label prediction for an input instance from the target domain represented by values for the set of features by applying the classifier to the input instance transformed using the stack of domain adaptation transform domains.

2. The device of claim 1 wherein some training instances are labeled with labels of a set of labels and the learning comprises:
performing supervised learning of the classifier on the training instances transformed using the stack of domain adaptation transform layers wherein the supervised learning optimizes the classifier respective to the labels of the labeled training instances.

3. The device of claim 1 wherein:
the set of features is a set of image features and the machine learning method includes the further operation of generating each training instance by extracting values for the set of image features from the corresponding image.

4. The device of claim 1 wherein the feature corruption probability vectors are nonuniform over the domains in accordance with a difference metric comparing the traffic light surveillance camera corresponding to a source domain with the traffic light surveillance camera corresponding to the target domain.

5. A device comprising:
a computer programmed to perform a machine learning method operating on training instances from a plurality of domains including one or more source domains and a target domain, with each training instance represented by values for a set of features wherein each domain corresponds to a traffic light surveillance camera and the training instances of each domain are images acquired by the corresponding surveillance camera of vehicles and the traffic light monitored by the corresponding surveillance camera, the set of labels indicate whether the imaged vehicle is running a red light, and the classifier outputs label predictions as to whether an imaged vehicle is running the traffic light, the machine learning method including the operations of:
performing domain adaptation using stacked marginalized denoising autoencoding (mSDA) operating on the training instances to generate a stack of domain adaptation transform layers wherein each iteration of the domain adaptation includes corrupting the training instances in accord with feature corruption probability vectors for the features that are non-uniform over the domains including:
corrupting the training instances from the one or more source domains in accord with a source domains feature corruption probability; and
corrupting the training instances from the target domain in accord with a target domain feature corruption probability that is different from the source domains feature corruption probability;

learning a classifier on the training instances transformed using the stack of domain adaptation transform layers; and
generating a label prediction for an input instance from the target domain represented by values for the set of features by applying the classifier to the input instance transformed using the stack of domain adaptation transform domains.

6. The device of claim 5 wherein some training instances are labeled with labels of a set of labels and the learning comprises:
performing supervised learning of the classifier on the training instances transformed using the stack of domain adaptation transform layers wherein the supervised learning optimizes the classifier respective to the labels of the labeled training instances.

7. The device of claim 5 wherein:
the set of features is a set of image features and the machine learning method includes the further operation of generating each training instance by extracting values for the set of image features from the corresponding image.

8. The device of claim 5 wherein the feature corruption probability vectors are nonuniform over the domains in accordance with a difference metric comparing the traffic light surveillance camera corresponding to a source domain with the traffic light surveillance camera corresponding to the target domain.

9. A non-transitory storage medium storing instructions executable by a computer to perform a machine learning method operating on training instances from a plurality of domains including (i) one or more source domains and (ii) a target domain, wherein each training instance represents a corresponding text-based document, and with each training instance represented by values for a set of features and some training instances being labeled with labels of a set of labels, the machine learning method including the operations of:
generating each training instance by extracting values for the set of features from the corresponding text-based document;
performing domain adaptation using stacked marginalized denoising autoencoding (mSDA) operating on the training instances with different feature corruption probabilities for different domains of the plurality of domains to generate a stack of domain adaptation transform layers operative to transform the domains to a common adapted domain, wherein the different feature corruption probabilities for different domains of the plurality of domains include one of:
(1) a source domains feature corruption probability used for training instances from the one or more source domains, and a target domain feature corruption probability used for training instances from the target domain, wherein the target domain feature corruption probability is different from the source domains feature corruption probability, or
(2) feature corruption probabilities p j , j=1, . . . , N and p t where N is the number of source domains with N being greater than or equal to two, p j is the feature corruption probability used for training instances from source domain j, and p t is the feature corruption probability used for training instances from the target domain, and wherein p j ≠p j′ for all j≠j′ and p t ≠p j for all j=1, . . . , N;

performing supervised or semi-supervised learning on the training instances transformed to the common adapted domain using the stack of domain adaptation transform layers to generate a document classifier that outputs label predictions from the set of labels for the training instances; and

Assignees

• Xerox Corp

Inventors

• Chidlovskii Boris
• Csurka Gabriela
• Clinchant Stéphane

Classifications

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/72

  Data preparation, e.g. statistical preprocessing of image or video features · CPC title

• G06V10/776

  Validation; Performance evaluation · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title