When Machine Learning Meets Philosophy: The Negative Ontology of the 'True Target'

A Disruptive Paper: Does the 'True Target' in Machine Learning Really Exist?

In the mainstream paradigm of machine learning, virtually all algorithms are built on a seemingly self-evident assumption — that a 'True Target' (TT) exists, and the model's mission is to approximate it as closely as possible. However, a newly published paper on arXiv, titled Negative Ontology of True Target for Machine Learning: Towards Evaluation and Learning under Democratic Supervision, launches a profound philosophical challenge against this foundational assumption.

By introducing the philosophical framework of 'negative ontology,' the study systematically questions the existence of the True Target and, on this basis, proposes an entirely new 'Democratic Supervision' evaluation and learning knowledge system, injecting unprecedented philosophical reflection into the theoretical foundations of machine learning.

Core Argument: From 'Existential Assumption' to 'Negative Ontology'

The current mainstream machine learning paradigm — whether supervised learning, semi-supervised learning, or reinforcement learning — fundamentally relies on a core premise: that behind the data lies some definite, approximable 'True Target.' Ground-truth labels in classification tasks, true values in regression tasks, and optimal policies in reinforcement learning are all concrete manifestations of this assumption.

However, the paper's authors point out that this 'existential assumption' does not hold in many real-world scenarios. In medical diagnosis, for example, different experts may give entirely different judgments on the same case; in judicial prediction, the so-called 'correct verdict' is itself fraught with controversy; and in the social sciences, many prediction targets simply do not have a uniquely determined 'true value.'

The paper explicitly adopts a 'negative ontology' stance, arguing at the philosophical level that in many important application scenarios, the True Target is not an objectively existing entity but rather a human construct. This shift in perspective may seem abstract, yet it has far-reaching implications for model evaluation, loss function design, and the entire learning paradigm.

Democratic Supervision: A New Framework for Learning and Evaluation

Based on the denial of the True Target's existence, the paper further proposes the innovative concept of 'Democratic Supervision.' Traditional supervised learning can be understood as a form of 'authoritative supervision' — where a single annotator or expert provides the sole 'correct answer,' and the model is trained to comply with this authority.

The Democratic Supervision framework proposes a fundamentally different approach:

• Elevating the ontological status of diverse annotations: Disagreements among multiple annotators are no longer treated as 'noise' but are acknowledged as reflecting the inherent plurality of the target itself.
• Reconstructing evaluation criteria: When the True Target does not exist, traditional metrics such as accuracy and mean squared error lose their theoretical grounding, necessitating a new evaluation philosophy.
• Democratizing learning objectives: The model no longer seeks to approximate a single 'true value' but instead learns to coordinate and integrate across multiple perspectives.

This framework essentially elevates machine learning from a purely technical optimization problem to a comprehensive issue involving epistemology and axiology.

Deeper Implications for Current AI Development

The significance of this paper extends far beyond academic discussion, touching on several key pain points in current AI development:

First, the philosophical roots of large model alignment. Current research on human alignment for large language models faces the challenge of 'whose values to align with.' The Democratic Supervision framework provides a theoretical tool for precisely this problem — if no single 'correct set of values' exists, then the goal of alignment should shift from approximating a fixed standard to seeking balance among diverse values.

Second, rethinking data annotation practices. The industry has long pursued 'high-quality annotations' and 'annotation consistency,' but from the perspective of negative ontology, forcibly pursuing consistency may actually be eliminating valuable information.

Third, a new foundation for AI fairness research. When we acknowledge that many prediction targets (such as credit scores and risk assessments) do not have objective 'true values,' fairness becomes not merely a matter of technical bias correction but requires fundamentally rethinking the purpose and meaning of prediction.

Academic Assessment and Future Outlook

From a methodological standpoint, the paper's greatest contribution lies in bridging the gap between analytic philosophy and machine learning, introducing ontological inquiry into a highly technical field. This interdisciplinary perspective is particularly valuable amid the current trend of increasingly 'engineering-oriented' AI research.

That said, the paper currently remains largely at the level of philosophical analysis and framework proposal. Concrete algorithmic implementations of Democratic Supervision and large-scale experimental validation still await future research. How to translate these profound philosophical insights into actionable technical solutions will be the greatest challenge facing this research direction.

Notably, as AI systems become increasingly involved in highly complex social domains such as healthcare, justice, and education, philosophical reflection on the 'True Target' will no longer be a speculative exercise confined to the academic ivory tower but a core issue concerning whether AI systems can be deployed responsibly. This paper sounds an important alarm: while pursuing ever-higher performance metrics, perhaps we need to pause and ask — what exactly are we optimizing for?