Are machines smart? Sentient? Conscious and cognizant? Leaving aside the slew of literature and movies from Space Odyssey to Short Circuit to Wall.E, questions of machine learning are pertinent in an age where we rely heavily on technology for almost every part of our day. The idea of a robotic vacuum cleaner smearing pet excrement around a house might be amusing (unless you're the owner of the pet and robotic vacuum cleaner in question) and an 'easily' repaired problem - how difficult is it to ensure you schedule the Roomba for a different time of day, or better yet, checking that the floor is clear of unpleasant surprises - but how does that extrapolate out to technology in far more important parts of society: to technology being used in health care or education or the justice system, to platforms and algorithms 'deciding' what and who will or won't be seen?
Broussard reminds us in this chapter that all of these so-called instances of learning have a common 'flaw': they are taught/designed/programmed by human beings, with all their inherent biases, ideologies, and ... ability to make sudden emotive decisions, either by necessity or whim. A computer may be able to predict who was likely to survive or die the sinking of the Titanic, all things - and data - being equal, but it cannot take into account what might happen if humans behave outside of the preset parameters. After all, the language used in these systems is 'highly precise'. As benign an example as predicting survival rates of an accident for which we already know the outcome might be, the reality is that these data sets have real impacts on real lives every day. Classifiers without context are inherently problematic - simply because context can and do change and these changes are usually driven by human behavior.
Another challenge,perhaps less obvious is the challenge of understanding the language required to 'teach' technology. In Broussard's chapter there are around 10 pages of numerical data that by Broussard's own admission are a challenge to process. By distancing ourselves from the ways in which technology 'learns' because they are too hard, we allow the data sets being designed to continue to be potentially either consciously or unconsciously oppressive, exclusionary, and dismissive. Creating a more diverse, more equal, and a more just digital, technological, and/or social landscape requires input not just from those who speak that mathematical, precise language but also from those who speak the nuanced, emotive non mathematical languages.
This was a good summary of the issue, and easy to read and understand: not only what you read, but your thoughts on the matter. Well done!
ReplyDelete