Beyond human-centered automated machine learning

Abstract

There has recently been a movement in AI research to go beyond ‘human-centered AI’, and to consider whether and how AI systems should promote non-anthropocentric values or values that go beyond human concerns. However, the research domain of automated machine learning (AutoML) lags behind on this shift. In some sense this is surprising, because AutoML has unusual potential to facilitate ‘beyond human AI’. For example, compared to typical supervised deep learning algorithms, AutoML algorithms can deal with much more complicated decision variables and objective functions that do not require a precise mathematical relation, making them much more flexible. This article makes a theoretical contribution by characterizing some of the ways in which AutoML could facilitate the creation of AI systems that go ‘beyond human’. First, it introduces AutoML and describes how the AutoML research community has recently investigated moving toward being human-centered in a particular sense. Next, the article describes how many other AI research communities, after moving toward human-centered AI in various senses in recent years, are now moving to consider the importance of ‘beyond human’ values. The article then turns to how AutoML research could possibly go ‘beyond human.’ Whereas many aspects of human-centered AutoML, such as transparency, may also advance ‘beyond human’ values, doing so in a reliable way would require moving beyond human-centered AutoML. The article then demonstrates that in comparison with typical supervised deep learning approaches, the flexibility of AutoML offers unique opportunities for facilitating ‘beyond human’ values that are not present in other areas of AI. If there is merit in developing systems that go beyond human values, as many people outside of AutoML seem to think, then AutoML offers a number of advantages.