What the edge tells the cloud: decentralized geographies of AI

Engineering notes

Key topics: self-driving car, self-driving, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Artificial Intelligence algorithms are increasingly deployed outside datacenters and at the so-called Edge, i.e., on end devices such as self-driving cars, smartphones, sensors, and drones, as well as on-premises servers in hospitals and other large but localized institutions. This matters, because, this paper argues, the materiality of the devices on which AI algorithms are deployed influences the “cognition” and behavior of those algorithms. The role of AI’s materiality has thus far not received the attention it deserves, and critical AI literature tends to consider AI algorithms as unproblematically individualized and spatially bounded entities, whereas an attention to materiality also shows how AI’s cognition is always decentralized and spatially distributed. One class of Edge AI algorithms is Federated Learning, where end devices learn partial models independently and sometimes collaborate to produce global models for inference and decision-making. This paper asks what forms of material and spatial cognitive configurations enable specific types of AI agency, and with what ethico-political consequences? Through close reading of ArXiv papers and patents, analyzed, combining Gilbert Simondon and N Katherine Hayles, showing how Federated Learning and Edge AI problematize existing understandings of AI algorithms as immaterial, centralized, and unified subjects. By showing how Edge AI and Federated Learning proceed by compressing models, orchestrating cognition across space, and federating models to produce global inferences, this paper shows that the materiality of the microchips and the spatiality of the data and feature spaces that Edge devices inhabit produce a specific, distributed, and contested form of algorithmic individuality . While distributed Machine Learning algorithms enable more diffuse and granular forms of surveillance and control, they also open the potential to inject partial and situated knowledge into Machine Learning, as an antidote to the current hubris.

Links and sources

• PDF from original source

Comments