The Compute Devide: The De-democratization of AI?

In a recent arXiv paper , analysis suggests that unequal access to computing capabilities may shift who and where AI is being innovated.

[…] we present systematic evidence that firms, in particular, large technology firms, are increasingly contributing more to AI research relative to other computer science areas. Our estimates suggest that Fortune 500 Global technology firms are publishing 44 additional papers annually per AI conference than the counterfactual. This is a significant change given that these firms’ average annual publication is only 23 papers per computer science conference. Similarly, elite universities (QS ranked 1-50) are publishing 40 additional papers per year per conference. In contrast, mid-tier universities (QS ranked 201-300) and lower tier universities (QS ranked 301-500) are publishing 14 and 5 fewer papers, respectively. Additionally, we document that Historically Black Colleges and Universities (HBCU), and Hispanic-serving institutions (Hispanic Association of College and Universities or HACU) are underrepresented in top AI venues.

Furthermore, using frequency-inverse document frequency or TF-IDF analysis, we provide evidence that the growing divergence in AI knowledge production between non-elite universities and large technology firms is attributable, in part, to the increasing divide in access to compute. We term this uneven distribution in access to computing power the “compute divide.” Our text analysis suggests that large technology firms are publishing more in deep learning areas than both elite and non-elite universities. […]

Ahmed, N., & Wahed, M. (2020). The De-democratization of AI: Deep Learning and the Compute Divide in Artificial Intelligence Research. ArXiv:2010.15581 [Cs].

Leave a Reply