Adapted domain specific class means classifier

Patent metadata

Field	Value
Publication number	US-10296846-B2
Application number	US-201514950544-A
Country	US
Kind code	B2
Filing date	Nov 24, 2015
Priority date	Nov 24, 2015
Publication date	May 21, 2019
Grant date	May 21, 2019

Abstract

A domain-adapted classification system and method are disclosed. The method includes mapping an input set of representations to generate an output set of representations, using a learned transformation. The input set of representations includes a set of target samples from a target domain. The input set also includes, for each of a plurality of source domains, a class representation for each of a plurality of classes. The class representations are representative of a respective set of source samples from the respective source domain labeled with a respective class. The output set of representations includes an adapted representation of each of the target samples and an adapted class representation for each of the classes for each of the source domains. A class label is predicted for at least one of the target samples based on the output set of representations and information based on the predicted class label is output.

First claim

What is claimed is:

1. A domain-adapted classification method comprising:
mapping an input set of representations to generate an output set of representations using a learned transformation, the input set of representations including a set of target samples from a target domain and, for each of a plurality of source domains, a class representation for each of a plurality of classes, the class representations each being representative of a set of source samples from the respective source domain labeled with a respective class, the output set of representations including an adapted representation of each of the target samples and an adapted class representation for each of the classes for each of the source domains;
predicting a class label for at least one of the target samples based on the output set of representations; and
outputting information based on the predicted class label,
wherein the mapping comprises, for at least one iteration:
learning a transformation that minimizes a reconstruction error when a corrupted set of representations, generated from the input set of representations, is transformed, with the transformation, to generate a reconstructed set of representations, and
outputting the reconstructed set of representations or adapted representations generated therefrom,

wherein each of the class representations and the target samples is a multidimensional representation comprising at least 10 dimensions, and
wherein at least one of the mapping of the input set of representations and the predicting of the class label is performed with a processor.

2. The method of claim 1 , wherein the class representations are class means.

3. The method of claim 1 , wherein the at least one iteration comprises a plurality of iterations and for a subsequent iteration, the input set is based on a reconstructed set of representations generated in a previous iteration.

4. The method of claim 3 , wherein the input set for the subsequent iteration is generated by performing a non-linear transformation on the reconstructed set of representations generated in the previous iteration.

5. The method of claim 1 , wherein the corrupted set of representations from the input set of representations corresponds to removing features from the input set with a predefined probability.

6. The method of claim 1 , wherein for at least one iteration, the learning of the transformation is performed without generation of the corrupted set of representations, by using a closed form-solution to marginalize out noise.

7. The method of claim 6 , wherein in the closed-form solution, the transformation
W= [P] [Q] −1 ,  (4),

where the expectation of Q for a given entry in matrix [Q], denoted

(

𝔼
⁡

[
Q
]

i
,
j

)

=

[

S
ij

⁢

q
i

⁢

q
j

,

if
⁢

⁢
i

≠
j

,

S
ij

⁢

q
i

,

if
⁢

⁢
i

=
j

,

]

and the expectation of P for a given entry in matrix [P], denoted [P] i,j =S ij q j ,
where i≠j indicates those values that are not on a diagonal of the matrix [P], and i=j those values that are on the diagonal,
q=[1−p, . . . ,1−p,1]∈R f+1 , where each element q i represents the probability of a feature i surviving the corruption, and q i q j represents the probability of features i and j both surviving the corruption=(1−p) 2 ;
p is a predefined probability;
f is a feature dimensionality, and
S=XX T is a covariance matrix of the uncorrupted data X and S ij is an element of the matrix S.

8. The method of claim 1 , wherein the predicting of the class label for the at least one of the target samples comprises computing one of:
a distance from the representation of the target sample in the output set to each of the adapted class representations in the output set; and
a distance from an augmented target representation to each of a set of augmented class representations, the augmented target representation being generated from the target representation in the output set, the augmented class representations being generated from the adapted class representations in the output set.

9. The method of claim 8 , wherein the augmented representations are generated by at least one of:
concatenating the output sets of a plurality of the iterations; and
concatenating the input set and the output set.

10. The method of claim 1 , wherein the predicting of the class label for the at least one of the target samples comprises, for each class, computing an optionally-weighted softmax distance from the adapted target representation to the adapted class representations for that class.

11. The method of claim 1 , wherein the mapping and the prediction are performed without access to the source samples.

12. The method of claim 1 , further comprising receiving the class representations for each of the plurality of source domains and the target samples and combining them to generate the input set of representations.

13. The method of claim 12 , wherein the combining comprises concatenating the class representations and the target samp

Assignees

• Xerox Corp

Inventors

• Csurka Gabriela
• Chidlovskii Boris
• Clinchant Stéphane

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N5/02

  Knowledge representation; Symbolic representation · CPC title

• G06N3/0454

  Physics · mapped topic

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title