Algorithmic biases are not (just) an algorithmic problem

There is a growing awareness of the biases that get thrown into the trained neural networks and other machine learning algorithms. Many of these biases are typically a direct result of the underlying discrimination and biases of how we sample and deal with the world. In more everyday terms, when we take pictures of some things (say people, cars, buildings and ‘scenes’) but tend to comparatively ignore other things (say random trash or pebbles on the pavement, or corners of buildings) those pictures serve as the material with which we end up training our ML algos. They will then, of course, preferentially learn about things we feed them more of, rather than things we feed them less of. There are now thoroughly publicized cases of these biases, so I’ll just mention two in passing to refresh your mind if you already know of them or to give you starting points for further reading if you did not know of this:

• Building and analyzing a word embedding model (e.g. word2vec or GloVe) on a large corpus of text (e.g. Wikipedia articles or collections of online articles) shows that ‘female’ and ‘woman’ are closer to words/concepts such as arts and humanities than ‘male’ or ‘man’, which tend to be more related to concepts and word contexts involving mathematics and engineering. While this is more or less a true statistical finding, we may or may not want the algorithms we use to make decisions based on this inherent bias in our current social environment
• A set of algorithms trained to judge beauty ‘objectively’ ended up choosing only white or almost-white people as the winners in this automated/ML-driven beauty contest. A large part of the reason was the disproportionately smaller amounts of non-white people who submitted photos — thus the input classes were unbalanced. This is a well known problem that affects most types of algorithms, but it is inherent in the social and cultural samplings that we undertake.

To overcome these biases, we can try to wrestle with the underlying mathematics and statistics in our models and take into account these biases. We can also work on eliminating some biases where we do not agree with them (perhaps certain gender stereotypes or the notion that certain races are more beautiful than others).

Ultimately, it is an amalgamation of these two, and other approaches, that we must make use of. If we don’t work on being more aware of our own biases in the world, then altering interpretations, decisions and actions in the world, based on these, will be hard. But we can’t eliminate all biases (nor would we want to — they are useful, statistically meaningful and without ‘biases’ we would be an amorphous blob of entropic mess) — so where some potentially undesirable ones exist and are hard to remove (due to biological or current socio-political and cultural constraints) we should develop better tools and methods in dealing with them.

How far our preferences propagate unchecked and without awareness into the world is the issue at hand. Increasingly capable algorithms can help us in checking and questioning our assumptions — but the algorithms and machines are not, by themselves, the solution to dealing with the inherent biases in the otherwise human world. We must become more aware and change ourselves too and not just our tools.