Methods and systems for identifying and reducing gender bias amplification
A multi-attribute bias amplification metric illustrates the need to consider multiple attributes when measuring bias amplification. For datasets that are perfectly balanced with respect to single attributes, bias amplification can still occur with respect to multi-attributes, regardless of whether raw or absolute differences are used. The metric can be used to show that methods used to mitigate single attribute bias can inadvertently increase multi-attribute bias amplification. Accordingly, the methods for determining bias amplification can provide a better understanding of the extent of bias a model introduces from training to prediction. Further, counterfactuals can be generated to decorrelate the co-occurrences of protected attributes with all background objects, both labeled and unlabeled. These generated counterfactuals can be used for both augmenting training and testing datasets. Using multilabel object classification, it can be demonstrated that training one counterfactual augmented training sets reduces bias along the protected attribute of perceived binary gender expression.
1. A method of determining an extent of a bias, in a model, across a plurality of group membership labels, comprising:
providing a training dataset having multiple training dataset attributes in a plurality of training images of the training dataset;
training the model to identify one of the plurality of group membership labels for a training image object in each of the plurality of training images, wherein the training associates the one of the plurality of group membership labels with multiple ones of the multiple training dataset attributes;
operating the trained model on a plurality of test images of a test dataset, each of the plurality of test images including a test image object and multiple test dataset attributes, to identify one of the group membership labels for each test image object based on the multiple test dataset attributes; and
determining a bias amplification in the identification of one of the plurality of group membership labels for each test image object in each of the plurality of test images.
2. The method of claim 1 , wherein the test image object and the training image object are images of people.
3. The method of claim 1 , wherein the bias is gender bias.
4. The method of claim 3 , wherein the plurality of group membership labels includes a male gender label and a female gender label.
5. The method of claim 1 , wherein the multiple training dataset attributes and the multiple test dataset attributes include actions describing the training image object and the test dataset object, respectively, of each of the plurality of training images and each of the plurality of test images, respectively.
6. The method of claim 1 , wherein the multiple training dataset attributes and the multiple test dataset attributes include physical objects within each of the plurality of training images and each of the plurality of test images, respectively.
7. The method of claim 1 , further comprising calculating a bias score correlated to the determined bias, wherein the bias score uses an absolute value of differences between the determined bias and a ground truth.
8. The method of claim 7 , wherein
the bias score of an attribute a with respect to group g is defined as:
b
(
a
,
g
)
=
𝒞
(
a
,
g
)
∑
g
′
∈
𝒢
𝒞
(
a
,
g
′
)
where ={g 1 , . . . g t } and ={a 1 , . . . a n } a set of t group membership labels and a set of n attributes, respectively, x∈ d and y=[g 1 , . . . g t , a 1 , . . . a n ]∈{0, 1} t+n are an image and ground truth labels, respectively, sampled from a dataset , g i ∈ denotes group membership and a i ∈ denotes the absence, a i =0, or presence, a i =1, of attribute i in a.
9. The method of claim 1 , wherein a bias amplification is reported based on a distance of the determined bias from a zero value.
10. The method of claim 1 , further comprising reporting a variance in the determined bias amplification, the variance signifying whether a bias amplification is uniform across all of the multiple attributes.
11. The method of claim 1 , further comprising reducing the bias amplification by generating synthetically balanced datasets based on at least one of the training dataset and the test dataset.
12. The method of claim 1 , further comprising generating a counterfactual image for each of the plurality of training images, wherein:
the training image object is an image of a person;
the bias is gender bias;
the plurality of group membership labels includes a male gender label and a female gender label; and
the counterfactual image changes a perceived binary gender expression, identified in the training dataset, of a first training image object from a male gender to a female gender, and the counterfactual image changes the perceived binary gender expression, identified in the training dataset, of a second training image object from the female gender to the male gender.
13. A method for measuring multi-attribute bias amplification to evaluate gender bias amplification, comprising:
providing a training dataset having multiple training dataset attributes in a plurality of training images of the training dataset;
training a model to identify one of a plurality of group membership labels for a person in each of the plurality of training images, wherein the training associates one of a male gender label or a female gender label with multiple ones of the multiple training dataset attributes;
operating the trained model on a plurality of test images of a test dataset, each of the plurality of test images including a test image person and multiple test dataset attributes, to identify one of the male gender label or the female gender label for each test image person based on the multiple test dataset attributes; and
determining the gender bias amplification in the identification of either the male gender label or the female gender label for each test image object in each of the plurality of test images.
14. The method of claim 13 , wherein the multiple training dataset attributes and the multiple test dataset attributes include at least one of actions, describing each of the plurality of training images and each of the plurality of test images, respectively, or physical objects within each of the plurality of training images and each of the plurality of test images, respectively.
15. The method of claim 13 , further comprising reducing the gender bias amplification by generating synthetically balanced datasets based on at least one of the training dataset and the test dataset.
16. The method of claim 13 , further comprising generating a counterfactual image for each of the plurality of training images, wherein the counterfactual image changes a perceived binary gender expression, identified in the training dataset, of a first training image object from a male gender to a female gender, and the counterfactual image changes the perceived binary gender expression, identified in a second training dataset, of the training image object from the female gender to the male gender.
17. 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 of improving computing efficiency of determining an extent of a bias, in a model, across a plurality of group membership labels with respect to multiple attributes, the method comprising:
providing a training dataset having multiple training dataset attributes in a plurality of training images of the training dataset;
training the model to identify one of the plurality of group membership labels for a training image object in each of the plurality of training images, wherein the training associates the one of the plurality of group membership labels with multiple ones of the multiple training dataset attributes;
operating the trained model on a plurality of test images of a test dataset, each of the plurality of test images including a test image object and multiple test dataset attributes, to identify one of the group membership labels for each test image object based on the multiple test dataset attributes; and
determining a bias amplification in the identification of one of the plurality of group membership labels for each test image object in each of the plurality of test images.
18. The non-transitory computer readable storage medium of claim 17 , wherein:
the test image object and the training image object are images of people;
the bias is gender bias; and
the plurality of group membership labels includes a male gender label and a female gender label.
19. The non-transitory computer readable storage medium of claim 17 , wherein the multiple training dataset attributes and the multiple test dataset attributes include at least one of actions, describing each of the plurality of training images and each of the plurality of test images, respectively, or physical objects within each of the plurality of training images and each of the plurality of test images, respectively.
20. The non-transitory computer readable storage medium of claim 17 , wherein the method further comprises reducing the bias amplification by generating synthetically balanced datasets based on at least one of the training dataset and the test dataset.