What measures have been taken to prevent discrimination or bias?
DeepVA applies data balancing, bias detection, and focal loss during training to ensure fair and non-discriminatory AI predictions.
Disclaimer:
This FAQ provides brief answers to frequently asked questions and serves as general guidance only. It does not replace legal advice or binding documentation. For the most up-to-date and legally relevant information, please refer to our official legal documentation at Deepva.ai/legal/.
If you have any legal questions or concerns, feel free to contact us directly.
Multiple measures are implemented to prevent discrimination or bias, addressing both data preparation and model training:
-
Data preparation:
Datasets are carefully analyzed, cleaned, and balanced to ensure that no social or demographic group is underrepresented. Biased patterns or content are actively identified and removed. -
Data distribution:
The standard deviation of variance is monitored to ensure an even distribution of data features. -
Model training:
During neural network training, focal loss functions are used to give more weight to hard-to-predict or underrepresented classes, helping reduce prediction bias.
These efforts aim to enable fair, balanced, and non-discriminatory decisions by the AI system.