Divinatory computation : the politics of Artificial Intelligence

Wednesday, 12 September, 2018 - 17:30

Faeeza Ballim (UJ), Keith Breckenridge (WISER), Richard Rottenburg (WISER) in conversation with Tshilidzi Marwala (VC of UJ)

WISER Seminar Room, 6th Floor Richard Ward

Nonconvex hyperspace distributionPanel Brief :  We aim in this panel to interrogate the meaning of the fourth industrial revolution and ask how transformative artificial intelligence models will be on the African continent. Machine learning practitioners have highlighted the increasing difficulties of introducing analytical rigour into these systems. At the Neural Information Processing Systems (NIPS) conference held in December 2017, Ali Rahimi argued that machine learning models more closely resembled alchemy than an exact science. The popular optimization technique gradient descent, which often involves millions of self-guided repetitions searching for a pathway to reduce errors in training inputs and results, also entails significant uncertainty about the absolute minimum of the cost function. We want to consider the significance of these arguments. How serious is the lack of confidence in machine learning techniques and how do practitioners deal with them? While the methods of machine learning are undeniably efficient, and capable of handling enormous datasets, they may also introduce arbitrariness and explanatory laziness. Machine learning holds out the ambiguous promise of addressing the informational weaknesses of African economies or potentially worsening them. Likewise while machine learning resources are strikingly open and accessible on this continent, especially in comparison with the older platforms for science and engineering training and research, they may endorse short-cut decision making and new asymmetries of knowledge production. In recent times, the availability of data in unprecedented amounts has improved the prospects of machine learning techniques, particularly of neural networks. While the conventional explanation is that data in Africa is notoriously hard to come by due to the lack of internet penetration, we consider the forms in which this data is available and how access to it is controlled. We also consider some of the sinister aspects of artificial intelligence models in its potential for reinforcing the power of bureaucratic authoritarian and illiberal states. Particularly concerning is the way that Chinese artificial intelligence companies have been easily accommodated in countries with weak privacy laws and with illiberal institutions.

Marwala Presentation

Ballim & Breckenridge Paper