Scientists from the University of California created a model for learning neural networks called Gender-Neutral Global Vectors (GN-GloVe). The development is intended for AI, specializing in the recognition of speech and texts. According to programmers, this training model will reduce the percentage of false gender associations. This is reported by Venture Beat.
Neural networks designed for speech recognition are trained on special data sets. However, these kits carry the imprint of a living language, filled with stereotypes. For example, the words "cook" or "secretary" are more often associated with the female sex, and "locksmith" or "welder" - with the male. Or, another examples: "doctor" is usually replaced by the pronoun "he", and "nurse" - "she".
Artificial intelligence, trained on such datasets, assimilates all the prejudices inherent in them. In particular, if a "doctor" is mentioned in the text without mentioning a particular sex, the neural network will more likely be considered a man. GN-GloVe, as claimed by its creators, removes the false associations with the sex.
This technology does not affect those areas where the sex is specified directly. To achieve this effect, the method determines gender-neutral words simultaneously with the formation of the semantic vectors of the text. Another advantage of development scientists call independence from the language being processed.
In a comparative analysis with GloVe, one of the most common teaching methods, the new model of scientists at the University of California showed 35% fewer mistakes due to false identification of a person's sex by type of activity.
Data sets for training contain many prerequisites for the formation of retraining errors. For example, smart speakers from Amazon and Google are 30% less likely to recognize English, pronounced with accent. And this problem is not just about speech: face recognition algorithms are worse at copying images of African Americans than Caucasians.
The bias of artificial intelligence bias surfaced in the work of Princeton University scientists in early 2017. While protection from such errors does not exist, however, similar GN-GloVe algorithms can in time reduce the bias error to an acceptable level.