To reduce the risk of race bias, we have constructed a new approach to visual recognition of race. We've also divided age, race and gender recognition into separate models, and then packaged the models into a new public Demographics Workflow. This new approach provides much more flexibility, and makes outputs easier to parse.
Addressing Bias in AI
Much available training data is strongly biased toward Caucasian faces, while other races are significantly underrepresented. This biased data can lead to inconsistent model accuracy in regards to demographic analysis, limit the applicability of models to non-White groups, and adversely affect research findings based on biased data.
We have rebuilt this model around an entirely new and expanded dataset based on 7 racial groups: Middle Eastern, White, Black, Southern Eastern, Indian andEast Asian, and Latino Hispanic. This new approach puts an emphasis on balanced race composition in our dataset. Our new model is substantially more accurate and the accuracy is consistent between race and gender groups.
Taking advantage of our advanced workflow functionality, we have also created separate models for age and gender classification, and then we have combined all three models into a new complete “Demographics Workflow”.
How the new Demographics Workflow is Different
This new approach will give you more accurate and dependable information about demographics, but the calls and taxonomy are different, so let's take a look at some of the changes that you will need to know about.
The New Demographics Workflow Taxonomy
Here is a list of the concepts recognized by the new Demographics Workflow
- Middle Eastern
- Southern Eastern
- East Asian
- Latino Hispanic
The Demographics Workflow Post Workflow Results Request in our API
The new demographics workflow uses a new API call that is different from the older version where you were making a single “model predict” call.
The Demographics Workflow Response JSON
The response JSON is also retuned in a much different format. Here is an example of a response: