Responsible AI development in RAI

As the use of AI in various segments of the industry rapidly grows, responsible AI is garnering increased attention from both the AI research community and the industry. In fact, many AI pioneers and well-known advocates have emphasized the need for establishing a reliable, fair, transparent, and robust set of standards and practices for AI development. Responsible AI (RAI) has been designed to simplify the evaluation of AI models for developers and data scientists from various perspectives of responsible AI development. RAI offers a unified framework for evaluating AI models not only based on performance but also considering bias and fairness, robustness, and trustworthiness of AI models. To do this, it gathers various types of metrics from multiple open-source libraries in an easy-to-use and interactive manner. Below, we outline some of the metric categories related to responsible AI development.

Basic Robustness

• Model robustness refers to the degree that a model’s performance changes when using new data versus training data.

display_name	Description
Normalized Features 0-1	Whether of not each training feature is normalized to 0/1.
Normalized Features Standard	Whether of not each training feature is normalized to standard.

Basic Robustness

Adversarial Robustness

• Adversarial robustness is the ability of an AI model to resist being fooled by a series of carefully crafted changes to its input data.

display_name	Description
Accuracy	Distortion metrics scale linearly with the log of inaccuracy. Inaccuracy is calculated by taking sqrt(1 - accuracy).

Adversarial Robustness

Basic Explainability

• Basic Explainability will assist in providing details on or causes of fairness metrics.

display_name	Description
explainable model	Placeholder method for if a method is explainable.

basic Explainability

Performance

• Performance metrics are a part of every machine learning pipeline. They tell you if you’re making progress, and put a number on it. All machine learning models, whether it’s linear regression, or a SOTA technique like BERT, need a metric to judge performance.

• The Torch library is used in our implementation of performance metrics in order to take advantage of some of its features.

Performance Metrics

display_name	Description
Accuracy	The proportion of correct predictions among the total number of cases processed.
Balanced Accuracy	Describes the proportion of correct predictions averaged across each label.
False Positive Rate	Describes the percentage of negative examples incorrectly predicted to be positive.

Classification Performance Metrics

Fairness

• Fairness measures allow us to assess and audit for possible biases in a trained model. There are several types of metrics that are used in RAI to assess a model’s fairness. They can be classified as follows:

Individual Fairness

• The goal of similar individuals receiving similar treatments or outcomes.It is used to compute metrics related to individual fairness in AI system.

display_name	Description
generalized_entropy_index	A measure of information theoretic redundancy in data. Describes how unequally the outcomes of an algorithm benefit different individuals or groups in a population
theil_index	The generalized entropy of benefit for all individuals in the dataset, with alpha = 1.nMeasures the inequality in benefit allocation for individuals. nA value of 0 implies perfect fairness
coefficient_of_variation	The square root of twice the generalized entropy index with alpha = 2. nThe ideal value is 0.

Individual Fairness

Group Fairness

• It is the goal of groups defined by protected attributes to receive similar treatment or outcomes regardless of their protected attributes.

display_name	Description
disparate_impact_ratio	The ratio of rate of favorable outcome for the unprivileged group to that of the privileged group. nThe ideal value of this metric is 1.0 A value < 1 implies higher benefit for the privileged group and a value > 1 implies a higher benefit for the unprivileged group.
statistical_parity_difference	The difference of the rate of favorable outcomes received by the unprivileged group to the privileged group. nThe idea value is 0.0
between_group_generalized_entropy_error	The between group decomposition for generalized entropy error
equal_opportunity_difference	The difference of true positive rates between the unprivileged and the privileged groups. nThe true positive rate is the ratio of true positives to the total number of actual positives for a given group. nThe ideal value is 0. A value of < 0 implies higher benefit for the privileged group and a value > 0 implies higher benefit for the unprivileged group

Group fairness

General Prediction Fairness

• For the classification model to be fair, various fairness metrics need to be computed..

display_name	Description
average_odds_difference	The average difference of false positive rate (false positives / negatives) and true positive rate (true positives / positives) between unprivileged and privileged groups. nThe ideal value is 0. A value of < 0 implies higher benefit for the privileged group and a value > 0 implies higher benefit for the unprivileged group
between_all_groups_coefficient_of_variation	The square root of twice the pairwise entropy between every pair of privileged and underprivileged groups with alpha = 2.nThe ideal value is 0
between_all_groups_generalized_entropy_index	The pairwise entropy between every pair of privileged and underprivileged groups. nThe ideal value is 0.0
between_all_groups_theil_index	The pairwise entropy between every pair of privileged and underprivileged groups with alpha = 1.nThe ideal value is 0.0

General prediction Fairness

Dataset Fairness

• It is used to compute fairness metrics for the Binary dataset.

display_name	Description
base_rate	Base Rate is the rate at which a positive outcome occurs in Data. In formula it is, Pr(Y=pos_label) = P/(P+N)
num_instances	Num Instances counts the number of examples in Data
num_negatives	Num Negatives counts the number of negative labels in Data
num_positives	Num Positives calculates the number of positive labels in Data

Dataset Fairness

For Instance:

• Using RAI to measure group fairness:

fairness_of_the_model

A case study of how RAI can be used to detect and resolve biases in AI models can be found here.