Undefinable True Target Learning

The inability to define the true target (TT) precisely in a TT learning task is a common challenge across various artificial intelligence (AI) application scenarios. In this article, we define this challenge as undefinable TT learning (UTTL). We explicitly propose that the fundamental assumption underlying UTTL is that the TT does not exist in the real world. To justify the necessity of introducing UTTL, we conducted a series of studies aimed at rigorously addressing the intrinsic question: why is UTTL needed? These investigations affirm that, under the assumption that the TT is nonexistent in the real world, UTTL is both necessary and significant. From the perspectives of problem definition, alternative formulations, methodological development, and application scenarios, we present a formal theoretical foundation for UTTL to effectively address learning tasks where the TT cannot be precisely defined. In doing so, this article not only establishes the theoretical basis of UTTL grounded in the explicitly stated assumption but also reveals, from a theoretical standpoint, the potential benefits of noisy labels in enabling UTTL.