Out-of-domain encoder training

Patent metadata

Field	Value
Publication number	US-11645514-B2
Application number	US-201916530457-A
Country	US
Kind code	B2
Filing date	Aug 2, 2019
Priority date	Aug 2, 2019
Publication date	May 9, 2023
Grant date	May 9, 2023

Abstract

A computer-implemented method includes using an embedding network to generate prototypical vectors. Each prototypical vector is based on a corresponding label associated with a first domain. The computer-implemented method also includes using the embedding network to generate an in-domain test vector based on at least one data sample from a particular label associated with the first domain and using the embedding network to generate an out-of-domain test vector based on at least one other data sample associated with a different domain. The computer-implemented method also includes comparing the prototypical vectors to the in-domain test vector to generate in-domain comparison values and comparing the prototypical vectors to the out-of-domain test vector to generate out-of-domain comparison values. The computer-implemented method also includes modifying, based on the in-domain comparison values and the out-of-domain comparison values, one or more parameters of the embedding network.

First claim

What is claimed is:

1. A computer-implemented method of training an encoder, the computer-implemented method comprising:
using an embedding network to generate prototypical vectors, each prototypical vector based on a corresponding label associated with a first domain;
using the embedding network to generate an in-domain test vector based on at least one data sample from a particular label associated with the first domain;
using the embedding network to generate an out-of-domain test vector based on at least one other data sample associated with a different domain;
comparing the prototypical vectors to the in-domain test vector to generate in-domain comparison values;
comparing the prototypical vectors to the out-of-domain test vector to generate out-of-domain comparison values; and
modifying, based on the in-domain comparison values and the out-of-domain comparison values, one or more parameters of the embedding network to generate one or more modified parameters for the embedding network.

2. The computer-implemented method of claim 1 , wherein modifying the one or more parameters of the embedding network comprises modifying one or more weights of the embedding network.

3. The computer-implemented method of claim 1 , further comprising determining a maximum likelihood of a true label based on the in-domain test vector, the in-domain test vector generated based on a training data sample selected from a label associated with the first domain.

4. The computer-implemented method of claim 3 , wherein the maximum likelihood of the true label is determined based on

L
in

=

expa
⁡
(

x
i
in

,

s

l
i

in

)

∑
l

expa
⁡
(

x
i
in

,

s

l
i

in

)

.

5. The computer-implemented method of claim 3 , wherein modifying the one or more parameters of the embedding network comprises:
selecting particular parameters that minimize a distance between the in-domain test vector and a particular prototypical vector associated with the maximum likelihood of the true label, wherein the particular parameters correspond to the one or more modified parameters.

6. The computer-implemented method of claim 5 , wherein minimizing the distance between the out-of-domain test vector and the particular prototypical vector is based on L gt =max[0, M 2 −max(F(x i in , S i in ))].

7. The computer-implemented method of claim 1 , wherein modifying the one or more parameters of the embedding network comprises:
selecting particular parameters that maximize distances between the out-of-domain test vector and the prototypical vectors, wherein the particular parameters correspond to the one or more modified parameters.

8. The computer-implemented method of claim 7 , wherein maximizing the distances between the out-of-domain test vector and the prototypical vectors is based on L ood =max[0, max(F(x j out , S l in )−M 1 )].

9. The computer-implemented method of claim 1 , further comprising:
randomly selecting a first group of one or more data samples from a first label associated with the first domain; and
randomly selecting a second group of one or more data samples from a second label associated with the first domain.

10. The computer-implemented method of claim 9 , further comprising:
encoding, using the embedding network, each data sample in the first group of one or more data samples to generate corresponding first sample vectors; and
encoding, using the embedding network, each data sample in the second group of one or more data samples to generate corresponding second sample vectors.

11. The computer-implemented method of claim 10 , further comprising:
performing an average-pooling operation on the first sample vectors to generate a first prototypical vector; and
performing the average-pooling operation on the second sample vectors to generate a second prototypical vector, wherein the prototypical vectors include at least the first prototypical vector and the second prototypical vector.

12. An apparatus comprising:
a processor; and
a memory coupled to the processor and storing instructions that, when executed by the processor, cause the processor to perform operations comprising:
using an embedding network to generate prototypical vectors, each prototypical vector based on a corresponding label associated with a first domain;
using the embedding network to generate an in-domain test vector based on at least one data sample from a particular label associated with the first domain;
using the embedding network to generate an out-of-domain test vector based on at least one other data sample associated with a different domain;
comparing the prototypical vectors to the in-domain test vector to generate in-domain comparison values;
comparing the prototypical vectors to the out-of-domain test vector to generate out-of-domain comparison values; and
modifying, based on the in-domain comparison values and the out-of-domain comparison values, one or more parameters of the embedding network to generate one or more modified parameters for the embedding network.

13. The apparatus of claim 12 , wherein modifying the one or more parameters of the embedding network comprises modifying one or more weights of the embedding network.

14. The apparatus of claim 12 , wherein the operations further comprise, for in-domain test data, performing an average pooling operation on each embedding per label to generate a respective prototypical vector.

15. The apparatus of claim 12 , wherein the operations further comprise determining a particular label that has a highest degree of similarity α(x i in , S l i in ) to the in-domain test vector.

16. The apparatus of claim 12 , wherein the operations further comprise classifying a particular test vector as “out-of-domain” if a similarity α(x i in , S l i in ) between the particular test vector and each prototypical vector is lower than a threshold.

17. A computer program product for training an encoder, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to c

Assignees

• IBM

Inventors

• Tan Ming
• Wang Dakuo
• Yu Mo
• Wang Haoyu
• Yu Yang
• Chang Shiyu
• Potdar Saloni

Classifications

• G06F17/16 Primary

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title