From Model-Centric to Data-Centric AI: A Paradigm Shift or Rather a Complementary Approach?

At its core, an artificial intelligence (AI) system mainly consists of an algorithm (code) that solves a problem by learning prototypical features from voluminous data clouds. Currently, there are two schools of thought on improving the performance of AI systems: model-centric AI and data-centric AI. In model-centric AI, developers of an AI system successively upgrade a devised model (algorithm/code), while holding the amount and type of data collected fixed. Conversely, practitioners of the data-centric AI keep the model fixed, while continuously improving the quality of data. Despite its dominance over the past three decades, model-centric AI has been recently criticised as being limited to businesses and industries where consumer platforms with hundreds of millions of users can freely rely on generalized solutions. This paper, rather than favoring any of the two approaches, supports a ‘both and’ position. We argue that overcoming the alleged limitation of model-centric AI may well require paying extra attention to the alternative data-centric approach. However, this shouldn't result in reducing interest in model-centric AI. We corroborate our position by the notion that successful ‘problem solving’ requires considering both the way we act upon things (algorithm) as well as harnessing knowledge of their states and properties (data).