CONCEPT
Confirmation Holism
Quine’s doctrine that the totality of human belief forms a web facing experience only at its edges—and the explanation of why model knowledge resists surgical editing, why catastrophic forgetting happens, and why no benchmark result settles a claim about a single belief.
The totality of our knowledge,
Willard Van Orman Quine wrote in 1951, forms “a man-made fabric which impinges on experience only along the edges.” No single statement confronts experience on its own; beliefs face the tribunal of evidence as a corporate body, and when a surprise strikes at the periphery, the disturbance propagates inward and we may revise almost anything to restore coherence. This is confirmation holism, and it is the first and firmest of the mappings between Quine’s philosophy and the practice of machine learning. A trained
neural network is, in its physical reality, this web rendered in arithmetic: knowledge distributed across millions of interlinked weights, no single weight encoding a retrievable fact, the whole adjusting together under the pressure of error. The geography Quine drew by hand is now an empirical property researchers can measure. Catastrophic forgetting—the disaster that strikes when fine-tuning on new data wrecks old competence—is precisely the disaster Quine’s