Game-theoretic frameworks for deep neural network rationalization

Patent metadata

Field	Value
Publication number	US-11657271-B2
Application number	US-201916658122-A
Country	US
Kind code	B2
Filing date	Oct 20, 2019
Priority date	Oct 20, 2019
Publication date	May 23, 2023
Grant date	May 23, 2023

Abstract

A method and system of determining an output label rationale are provided. A first generator receives a first class of data and selects one or more input features from the first class of data. A first predictor receives the one or more selected input features from the first generator and predicts a first output label. A second generator receives a second class of data and selects one or more input features from the second class of data. A second predictor receives the one or more selected input features from the second generator and predicts a second output label. A discriminator receives the first and second output labels and determines whether the selected one or more input features from the first class of data or the selected features of the one or more input features from the second class of data, more accurately represents the first output label.

First claim

What is claimed is:

1. A computer implemented class-wise adversarial rationalization (CAR) system for determining an output label rationale, comprising:
a first generator network operative to receive a first class of data and select one or more input features from the first class of data;
a first predictor network operative to receive the one or more selected input features from the first generator network and predict a first output label based on the received one or more selected input features from the first generator network;
a second generator network operative to receive a second class of data and select one or more input features from the second class of data;
a second predictor network operative to receive the one or more selected input features from the second generator network and predict a second output label based on the received one or more selected input features from the second generator network; and
a discriminator network operative to receive the first and second output labels and to determine whether the selected one or more input features from the first class of data or the selected features of the one or more input features from the second class of data, more accurately represents the first output label.

2. The system of claim 1 , wherein the first class of data represents a first rationale (R 1 ) consistent with the first output label.

3. The system of claim 1 , wherein the second class of data represents a second rationale (R 2 ) consistent with the second output label.

4. The system of claim 1 , wherein the second output label is a complement of the first output label.

5. The system of claim 1 , wherein the first and second class of data are each in natural language.

6. The system of claim 1 , wherein:
the first generator network is trained on a subset of the first class of data;
the second generator network is trained on a subset of the second class of data; and
the first output label is inconsistent with the output label of the second output label.

7. The system of claim 1 , wherein the first generator network is configured to play a cooperative game with the discriminator network.

8. The system of claim 7 , wherein the second generator network is configured to play an adversarial game with the discriminator network.

9. The system of claim 8 , wherein the cooperative game between the first generator network and the discriminator network maximizes a predictive accuracy of the one or more input features selected by the first generator network to predict the first output label.

10. The system of claim 8 , wherein the adversarial game between the second generator network and the discriminator network is operative for the second generator network to attempt to convince the discriminator network that the features selected by the second generator network from the second class of data are consistent with the first output label.

11. The system of claim 1 , wherein the discriminator network is operative to provide a scaled result based on the selected one or more input features from the first class of data and the selected one or more input features from the second class of data.

12. The system of claim 1 , wherein the first generator network and the second generator network are each configured to convince the discriminator network that that they are factual generators for the first output label.

13. A computing device comprising:
a processor;
a storage device coupled to the processor;
a program stored in the storage device, wherein an execution of the program by the processor configures the computing device to perform acts, comprising:
receiving a first class of data by a first generator network;
selecting one or more input features from the first class of data by the generator network;
receiving, by a first predictor network, the one or more selected input features from the first generator network;
predicting, by the first predictor network, a first output label based on the received one or more selected input features from the first generator network;
receiving, by a second generator network, a second class of data;
selecting, by the second generator network, one or more input features from the second class of data;
receiving, by a second predictor network, the one or more selected input features from the second generator network;
predicting, by the second predictor network, a second output label based on the received one or more selected input features from the second generator network;
receiving, by a discriminator network, the first and second output labels; and
determining, by the discriminator network, whether the selected one or more input features from the first class of data or the selected features of the one or more input features from the second class of data, more accurately represents the first output label.

14. The computing device of claim 13 , wherein:
the first class of data represents a first rationale (R 1 ) consistent with the first output label; and
the second class of data represents a second rationale (R 2 ) consistent with the second output label.

15. The computing device of claim 13 , wherein execution of the program by the processor further configures the computing device to perform acts, comprising:
training the first generator network on a subset of the first class of data; and
training the second generator network on a subset of the second class of data, wherein the first output label is inconsistent with the output label of the second output label.

16. The computing device of claim 13 , wherein execution of the program by the processor further configures the computing device to perform acts, comprising: playing a cooperative game between the first generator network and the discriminator network.

17. The computing device of claim 16 , wherein execution of the program by the processor further configures the computing device to perform acts, comprising: playing an adversarial game between the second generator network and the discriminator network.

18. The computing device of claim 17 , wherein the cooperative game between the first generator network and the discriminator network maximizes a predictive accuracy of the one or more input features selected by the first generator network to predict the first output label.

19. The computing device of claim 13 , wherein the first generator network and the second generator network are each configured to convince the discriminator network that that they are factual generators for the first output label.

20. A non-transitory computer readable storage medium tangibly embodying a computer readable program code having computer readable instructions that, when executed, causes a computer device to carry out a method, comprising:
receiving a first class of data by a first generator network;
selecting one or more input features from the first class of data by the generator network;
receiving, by a first predictor network, the one or more selected input features from the first generator network;
predicting, by the first predictor network, a first output label based on the received one or more selected input features from the first generator network;
receiving, by a second generator network, a second class of data;
selecting, by the second generator network, one or more input features from the second class of data;
receiving, by a second predictor network, the one or more selected input features from the second generator network;
predicting, by the second predictor network, a second output label based

Assignees

• IBM
• Massachusetts Inst Technology

Inventors

• Chang Shiyu
• Yu Mo
• Zhang Yang
• Jaakkola Tommi S

Classifications

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/094

  Adversarial learning · CPC title