Hi All,
I would be very much interested in any exchange of best practices/ideas/initiatives on addressing bias in AI systems (in particular gender bias).
P.s. Great initiative by the Commission.
Thank You!
Hi,
This is a very specific case example, but I'm sending this because from your question is not clear whether you look for generic info, or specific techniques & algorithms.
Check this one - the video about Debiasing word embeddings may give you some insights.
https://www.coursera.org/learn/nlp-sequence-models/home/week/2
Cheers,
Jarek
Hi Emmanouil,
I'm sending you a link to a position paper I presented at the ICSE conference this year on gender bias in AI - hope this helps.
Thanks,
Susan.
Hi Emmanouil,
I'm slightly puzzled by the context of your question, and judging by the difference responses provided, I'm not the only one.
Nevertheless, from a strictly data scientific poin of view - you want your data to be as bias-free as possible (not to confuse with the model bias, as in the variance - bias trade-off). If the data to be modelled were to be biased, and you attempt to perform some sort of multicategorical classification, having the data biased may result in the algorithm learning to detect only the dominant category, rather than trully determine the correct category. Not all estimators suffer from this sensitivity issue (decision trees-based techniques do pretty well on imbalanced data), but in principle you should start by considering the three options: downsampling the dominant category i.e. removing some samples of the most common class from the dataset; upsample the underrepresented classes, making sure you don't alter the true nature of the real data; mix the two approaches, typically in the order of upsampling followed by downsampling.
Not sure this is the angle of perspective you were aiming at, but if - hope this helps to guide you in the right direction.
Regards,
Paweł
From a policy making/governmental perspective, I think it's important to have a diverse group of people working on AI. It should not only be people of a certain age group and gender, but a cross-section of society working on AI-related policy issues (to the extent this is possible in a government administration).
When we put together our AI expert group, we also made sure that it is gender and geographically balanced to the extent possible.
Hi,
regarding bias...or neutrality please review the TEDx talk I gave last month in Rome. The audio/slides/text are available here: https://www.researchgate.net/project/NeuralMimicry
The video is expected to be available on TEDx/Youtube soon.
Please contact me directly if you would like to discuss further.
Paul Isaac's
I can recommend Ng's deep learning specialization offered by Coursera, specifically the lecture about de-biasing word embeddings. You can audit the lecture for free : https://www.coursera.org/learn/nlp-sequence-models/lecture/zHASj/debiasing-word-embeddings
Word embeddings or encodings are used a lot in natural language processing / understanding tasks.
Most word embeddings (or encodings) from available corpus are biased because the corpus themselves are biased. So it is frequent to obtain analogies like Father: Doctor as Mother: Nurse. There are some techniques you can apply to remove these biases and make your bots, personal assistants, etc. gender unbiased. The referenced lecture gives an example of such techniques.
Hi Emmanouil,
Take a look at the recent work of the EU Fundamental Rights Agency on this topic, namely "#BigData: Discrimination in data-supported decision making (May 2018)".It suggests 4 potential ways of minimising the risk of discrimination in using big data for automated decision making.
I will be based in Ghent next year doing research (funded by the EU) on this question. We could have a chat then.
K.S.
Yes, that would be great - my schedule is more flexible next month if you like?
Hi Katerina, I would love to get in touch with you as I am also working on this topic, and I am from Ghent. Maybe we could catch up? Would be interested in which department you are going to do this research. Thanks for getting in touch with me.
I believe that equal to a human organization, an AI organization could be divided into two or more basic operational models. For example, a significant difference in what capabilities an AI might have could result in the illusion of separation between the new species, exactly the way we differentiate male and female beings based on significant external and operational differences, even when acknowledging that we are both humans.
