Ricardo Baeza-Yates from Institute for Experiential AI at Northeastern University gave a talk titled “Ethics in AI: A Challenging Task”. Machine-learning algorithms are vulnerable to data biases that are often amplified and lead to unfair decision-making. For example, a system used by the New York State Department of Justice to determine bail amounts is negatively biased toward people of color. Yet, crucially, that bias has little to do with the algorithm itself, but rather how it is trained. The algorithm is trained upon judges’ decisions, which are racially biased. What's better? A biased algorithm? Or a noisy judge? A biased algorithm will produce the same outcome for two very similar cases, whereas a noisy judge (e.g., research showed that judges tend to be harsher before lunch) might well decide on two different sentences for the same cases. Unlike the algorithm, the judge is not deterministic.

Furthermore, beyond the problem of data bias, even in the presence of a perfect training dataset, a machine-learning expert should still consider that the algorithm has been:

• Trained with data that does not capture the entire context of the problem.

• Optimized for accuracy, which might not always be important. Arguably, it
would be more important to measure the impact of misclassifications,
however rare they may be. In the medical domain, where misclassifications
have significant consequences, measuring their impact is far more important
than measuring accuracy per se.

• Designed to produce deterministic verifiable outputs. Yet, in circumstances
of low confidence, rather than a clear-cut answer, it could be less harmful
to output “I do not know.”

To find out more, see the Nokia Bell Labs Responsible AI hub: [ Ссылка ]

свернуть